Skip to content

armgit / ODS

Public Imported from GitHub
0

ODS V3 Pre-Release: Public testing and refinement ahead of the official V3 launch. Turn your PC, Mac, or Linux box into a private AI server.

182 KB
Raw Blame
{
  "version": 2,
  "models": [
    {
      "id": "qwen3.5-2b-q4",
      "name": "Qwen 3.5 2B",
      "family": "qwen",
      "gguf_file": "Qwen3.5-2B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3.5-2B-GGUF/resolve/f6d5376be1edb4d416d56da11e5397a961aca8ae/Qwen3.5-2B-Q4_K_M.gguf",
      "gguf_sha256": "aaf42c8b7c3cab2bf3d69c355048d4a0ee9973d48f16c731c0520ee914699223",
      "size_bytes": 1280835840,
      "size_mb": 1281,
      "vram_required_gb": 3,
      "context_length": 65536,
      "max_context_length": 262144,
      "block_count": 24,
      "attention_layer_count": 6,
      "attention_head_count_kv": 2,
      "attention_key_length": 256,
      "attention_value_length": 256,
      "recurrent_state_bytes": 20201472,
      "architecture_source": {
        "repo": "Qwen/Qwen3.5-2B",
        "revision": "15852e8c16360a2fea060d615a32b45270f8a8fc",
        "file": "config.json"
      },
      "selection": {
        "discrete": 20,
        "unified": 20,
        "cpu": 20
      },
      "quantization": "Q4_K_M",
      "specialty": "Low-VRAM Chat",
      "description": "Qwen's compact hybrid model with a native 256K context limit, tool-use support, broad multilingual coverage, and a 64K operating target. Pixel supports it in adaptive mode; historical cross-app probes show lower reliability on four of six fleet hosts, so complex work may need more guidance or a stronger model.",
      "tokens_per_sec_estimate": 120,
      "llm_model_name": "qwen3.5-2b",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "unsupported_until_revalidated",
          "label": "App chat needs revalidation",
          "reason": "Fresh exact-commit six-host validation loaded Qwen 3.5 2B and passed ODS Talk everywhere, but LiteLLM and Open WebUI exact-token probes emitted truncated reasoning instead of the requested answer on strixy.",
          "evidence": "fleet-test/runs/2026-07-27T08-54-24Z-model-ui-product-b5da3792c281-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T09:13:58Z",
          "productSha": "b5da3792c281e0ba8f679e33876ee3de902a7dd6",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129"
        },
        "hermes_talk": {
          "status": "verified",
          "label": "ODS Talk verified across fleet",
          "reason": "The same exact-commit run loaded Qwen 3.5 2B at 64K context and passed challenge-bound ODS Talk on all six Linux, macOS, and Windows hosts before app-consumer validation.",
          "evidence": "fleet-test/runs/2026-07-27T08-54-24Z-model-ui-product-b5da3792c281-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T09:13:58Z",
          "productSha": "b5da3792c281e0ba8f679e33876ee3de902a7dd6",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "perplexica": {
          "status": "unsupported_until_revalidated",
          "label": "Perplexica needs revalidation",
          "reason": "Perplexica exact-token probes were instruction-following failures on spark, m5-mbp, and windows-laptop after successful candidate load and Talk inference. The model refused or omitted the required verification phrase while the services remained healthy.",
          "evidence": "fleet-test/runs/2026-07-27T08-54-24Z-model-ui-product-b5da3792c281-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{spark,m5-mbp,windows-laptop}",
          "recordedAt": "2026-07-27T09:13:58Z",
          "productSha": "b5da3792c281e0ba8f679e33876ee3de902a7dd6",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129"
        },
        "agent_viability": {
          "status": "not_agent_viable",
          "label": "Adaptive; reliability varies",
          "reason": "Only tower2 and strix-halo completed the full rotation. Four hosts failed required candidate-phase app probes, although every failure recovery restored the original model and deleted the test artifact cleanly.",
          "evidence": "fleet-test/runs/2026-07-27T08-54-24Z-model-ui-product-b5da3792c281-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T09:13:58Z",
          "productSha": "b5da3792c281e0ba8f679e33876ee3de902a7dd6",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129"
        }
      },
      "runtime_profiles": [
        {
          "id": "cpu-64k-q8-kv",
          "label": "CPU 64K context with Q8 KV cache and capped checkpoints",
          "backend": "cpu",
          "context_length": 65536,
          "fit_label": "64K CPU Q8 KV profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0",
            "LLAMA_ARG_CTX_CHECKPOINTS": "4",
            "LLAMA_ARG_CACHE_RAM": "512",
            "LLAMA_SERVER_MEMORY_LIMIT": "4G"
          }
        }
      ]
    },
    {
      "id": "jamba-reasoning-3b-q4",
      "name": "AI21 Jamba Reasoning 3B",
      "family": "jamba",
      "gguf_file": "jamba-reasoning-3b-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ai21labs/AI21-Jamba-Reasoning-3B-GGUF/resolve/462e08a43c3c32f6b8b85f79ff0796e484d7b65a/jamba-reasoning-3b-Q4_K_M.gguf",
      "gguf_sha256": "5c8edf36ec3ad9792a639db8d6865e479038226cf8fc71ef47331c611854f6c8",
      "size_bytes": 1932698048,
      "size_mb": 1933,
      "vram_required_gb": 3,
      "context_length": 65536,
      "max_context_length": 262144,
      "quantization": "Q4_K_M",
      "specialty": "Efficient Reasoning",
      "description": "AI21's compact hybrid reasoning model combines 26 Mamba layers with 2 attention layers for low-memory long-context work. Pixel supports it in adaptive mode; historical OpenCode probes show lower reliability on two independent fleet backends, so complex work may need more guidance or a stronger model.",
      "tokens_per_sec_estimate": 90,
      "llm_model_name": "jamba-reasoning-3b",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "verified",
          "label": "App chat verified across fleet",
          "reason": "Fresh exact-commit six-host validation loaded the pinned Jamba artifact and passed the required LiteLLM and Open WebUI exact-token probes everywhere while route evidence named the live Jamba model.",
          "evidence": "fleet-test/runs/2026-07-27T10-47-19Z-model-ui-product-ca730791caca-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T11:12:30Z",
          "productSha": "ca730791cacaadb0280f63f3fc9f8b8ef70e4ebb",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "hermes_talk": {
          "status": "verified",
          "label": "ODS Talk verified across fleet",
          "reason": "ODS Talk returned its challenge-bound verification phrase after the live Jamba swap on all six hosts, including both Windows Lemonade machines.",
          "evidence": "fleet-test/runs/2026-07-27T10-47-19Z-model-ui-product-ca730791caca-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T11:12:30Z",
          "productSha": "ca730791cacaadb0280f63f3fc9f8b8ef70e4ebb",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "opencode": {
          "status": "unsupported_until_revalidated",
          "label": "OpenCode needs revalidation",
          "reason": "OpenCode routed to the exact live Jamba model but returned unrelated or repetitive reasoning instead of its requested verification phrase on tower2 through Linux llama-server and strix-halo through Lemonade. Four other hosts passed the same probe.",
          "evidence": "fleet-test/runs/2026-07-27T10-47-19Z-model-ui-product-ca730791caca-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo}",
          "recordedAt": "2026-07-27T11:12:30Z",
          "productSha": "ca730791cacaadb0280f63f3fc9f8b8ef70e4ebb",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo"],
          "expiresAt": "2026-08-27T11:12:30Z"
        },
        "agent_viability": {
          "status": "not_agent_viable",
          "label": "Adaptive; reliability varies",
          "reason": "Four hosts completed the full rotation, but two independent backends failed the required OpenCode candidate probe after successful load and Talk inference. Every host restored its original model and deleted the test artifact cleanly.",
          "evidence": "fleet-test/runs/2026-07-27T10-47-19Z-model-ui-product-ca730791caca-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T11:12:30Z",
          "productSha": "ca730791cacaadb0280f63f3fc9f8b8ef70e4ebb",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "phi4-mini-q4",
      "name": "Phi-4 Mini",
      "family": "phi",
      "gguf_file": "Phi-4-mini-instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Phi-4-mini-instruct-GGUF/resolve/78eb92a46fc37e6b524df991ed9aca9bc6aa7b80/Phi-4-mini-instruct-Q4_K_M.gguf",
      "gguf_sha256": "88c00229914083cd112853aab84ed51b87bdf6b9ce42f532d8c85c7c63b1730a",
      "size_bytes": 2491874272,
      "size_mb": 2490,
      "vram_required_gb": 4,
      "context_length": 128000,
      "block_count": 32,
      "embedding_length": 3072,
      "attention_head_count": 24,
      "attention_head_count_kv": 8,
      "install_recommendation": false,
      "install_recommendation_reason": "Needs 128 KiB of KV cache per token, so only 8-16K context fits on 4-8 GB, below the 16K Pixel and 64K Hermes floors. Kept for manual activation (the Apple 8 GB profile still applies).",
      "quantization": "Q4_K_M",
      "specialty": "Balanced",
      "description": "Microsoft's compact model with 128K context. Punches above its weight on reasoning and code.",
      "tokens_per_sec_estimate": 130,
      "llm_model_name": "phi-4-mini",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "verified",
          "reason": "Fleet diagnostic on m5-mbp at 42b3a95c returned the requested verification code in 4/4 direct OpenAI-compatible chat completions."
        },
        "hermes_talk": {
          "status": "verified",
          "label": "ODS Talk revalidated on Strixy",
          "reason": "Challenge-bound ODS Talk passed again on Strixy with Phi-4 Mini during the exact e4fd1cae six-cycle qualification; this proves Talk on this host, not all-app, Pixel, or Open WebUI compatibility.",
          "evidence": "ods-public-beta-fleet-green-20260911/release-strixy-e4fd1cae-h3043027-r401/model-ui/cycle-005/strixy-wsl-beta/model-ui.json",
          "recordedAt": "2026-09-21T07:30:37Z",
          "productSha": "e4fd1cae6232af0a9d7438317002b55f94bea2bc",
          "harnessSha": "30430271ba0b97c83056272a4f4877cb65dee122",
          "hostScope": ["strixy"],
          "expiresAt": "2026-10-21T00:00:00Z"
        },
        "open_webui": {
          "status": "unsupported_until_revalidated",
          "label": "Open WebUI revalidation required on Strixy",
          "reason": "The exact e4fd1cae Strixy revalidation loaded Phi-4 Mini through native Lemonade, proved the signed expected route and returned HTTP 200, but Open WebUI mutated a verification digit by changing the requested token from ODSVAL-A6B4422ACC-461A91E67B to ODSVAL-A6B4422ACC-461A891B7B. Keep this host/model/app pairing out of all-app release coverage until exact-answer behavior is revalidated.",
          "evidence": "ods-public-beta-fleet-green-20260911/release-strixy-e4fd1cae-h3043027-r401/model-ui/cycle-005/strixy-wsl-beta/model-ui.json",
          "recordedAt": "2026-09-21T07:30:37Z",
          "productSha": "e4fd1cae6232af0a9d7438317002b55f94bea2bc",
          "harnessSha": "30430271ba0b97c83056272a4f4877cb65dee122",
          "hostScope": ["strixy"],
          "expiresAt": "2026-10-21T00:00:00Z"
        },
        "pixel_agent": {
          "status": "unsupported_until_revalidated",
          "label": "Pixel revalidation required on Strixy",
          "reason": "The exact e4fd1cae Strixy revalidation reached Phi-4 Mini over the signed expected route with HTTP 200 and completed retained Pixel SSE, but dropped the leading O from the requested phrase. This repeats earlier exact-answer failures despite correct routing, so keep the host/model pairing excluded until response behavior is revalidated.",
          "evidence": "ods-public-beta-fleet-green-20260911/release-strixy-e4fd1cae-h3043027-r401/model-ui/cycle-005/strixy-wsl-beta/model-ui.json",
          "recordedAt": "2026-09-21T07:30:37Z",
          "productSha": "e4fd1cae6232af0a9d7438317002b55f94bea2bc",
          "harnessSha": "30430271ba0b97c83056272a4f4877cb65dee122",
          "hostScope": ["strixy"],
          "expiresAt": "2026-10-21T00:00:00Z"
        },
        "agent_viability": {
          "status": "unsupported_until_revalidated",
          "label": "All-app revalidation required on Strixy",
          "reason": "The exact e4fd1cae Strixy revalidation did not sustain all-app coverage for Phi-4 Mini: Pixel dropped the first challenge character and Open WebUI mutated a verification digit despite exact signed routes and HTTP 200 responses. Talk, LiteLLM, OpenCode, and Perplexica passed, and failure recovery restored the Qwen 3.6 baseline.",
          "evidence": "ods-public-beta-fleet-green-20260911/release-strixy-e4fd1cae-h3043027-r401/model-ui/cycle-005/strixy-wsl-beta/model-ui.json",
          "recordedAt": "2026-09-21T07:30:37Z",
          "productSha": "e4fd1cae6232af0a9d7438317002b55f94bea2bc",
          "harnessSha": "30430271ba0b97c83056272a4f4877cb65dee122",
          "hostScope": ["strixy"],
          "expiresAt": "2026-10-21T00:00:00Z"
        }
      },
      "runtime_profiles": [
        {
          "id": "apple-silicon-8gb-16k",
          "label": "Apple Silicon 8GB 16K context with Q4 KV cache",
          "backend": "apple",
          "host_arch": [
            "arm64"
          ],
          "memory_type": "unified",
          "system_ram_min_gb": 7,
          "system_ram_max_gb": 10,
          "context_length": 16384,
          "estimated_required_gb": 4.35,
          "estimate_source": "fleet-validated",
          "fit_label": "16K 8GB Q4 KV profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "auto",
            "LLAMA_ARG_CACHE_TYPE_K": "q4_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q4_0"
          }
        }
      ]
    },
    {
      "id": "phi3.5-mini-q4",
      "name": "Phi-3.5 Mini",
      "family": "phi",
      "gguf_file": "Phi-3.5-mini-instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/bartowski/Phi-3.5-mini-instruct-GGUF/resolve/6d70da17e749a471ccb62ade694486011a75cda3/Phi-3.5-mini-instruct-Q4_K_M.gguf",
      "gguf_sha256": "e4165e3a71af97f1b4820da61079826d8752a2088e313af0c7d346796c38eff5",
      "size_bytes": 2393232672,
      "size_mb": 2393,
      "vram_required_gb": 4,
      "context_length": 128000,
      "quantization": "Q4_K_M",
      "specialty": "Long Context",
      "description": "Microsoft's previous-generation compact Phi model with 128K context. Useful as a lightweight long-context fallback.",
      "tokens_per_sec_estimate": 125,
      "llm_model_name": "phi-3.5-mini-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "unsupported_until_revalidated",
          "label": "Windows 8GB runtime unsupported",
          "reason": "Fresh exact-commit Windows revalidation loaded Phi-3.5 Mini at the required 65,536-token context with Q4 KV cache, but llama.cpp projected 9,344 MiB of device memory against 6,912 MiB free and challenge-bound direct probes decoded at only 0.47-0.49 tok/s before the Hermes prompt. Keep this model out of Windows 8GB release coverage until a viable partial-offload or lower-memory runtime is implemented.",
          "evidence": "fleet-test/runs/2026-07-24T07-04-00Z-windows-profile-candidates-route-validation/cycle-002/windows-laptop",
          "recordedAt": "2026-07-24T07:16:00Z",
          "productSha": "011bbabd74b72476862ad9a741dc4876b57e9521",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "perplexica": {
          "status": "unsupported_until_revalidated",
          "label": "Perplexica orchestration unsupported",
          "reason": "Fresh exact-commit release cycles loaded Phi-3.5 Mini through the browser UI on tower2 (Linux llama-server) and strix-halo (Lemonade), then passed challenge-bound ODS Talk, LiteLLM, Open WebUI, and OpenCode where deployed. Perplexica alone failed to complete within 180 seconds on both independent backends, while its baseline probes and the following Granite H-Micro cycle passed. Keep Phi-3.5 Mini out of all-app release coverage until Perplexica orchestration is revalidated.",
          "evidence": "fleet-test/runs/2026-07-24T09-22-58Z-release-product-4cd022b82609-harness-954deb755b07-hosts-six-scope-6cycle-switchboard/model-ui/cycle-001/{tower2,strix-halo}",
          "recordedAt": "2026-07-24T10:09:00Z",
          "productSha": "4cd022b826093913dda7137b2dd69502628d7f50",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "globalScope": true,
          "expiresAt": "2026-08-24T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "phi4-mini-reasoning-q4",
      "name": "Phi-4 Mini Reasoning",
      "family": "phi",
      "gguf_file": "Phi-4-mini-reasoning-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/SandLogicTechnologies/Phi-4-mini-reasoning-GGUF/resolve/40e6071714bc133bd46ad2dd62eaca1b288219a1/Phi-4-mini-reasoning-Q4_K_M.gguf",
      "gguf_sha256": "81878401a2f8160473649af89560a7fc0932f3623e4f6e58143d5dcbf71d6480",
      "size_bytes": 2491875232,
      "size_mb": 2490,
      "vram_required_gb": 4,
      "context_length": 128000,
      "quantization": "Q4_K_M",
      "specialty": "Reasoning",
      "description": "Microsoft's compact Phi reasoning model with 128K context. A low-VRAM candidate for math-heavy and logic-heavy chat workflows.",
      "tokens_per_sec_estimate": 110,
      "llm_model_name": "phi-4-mini-reasoning",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-17T00-45Z on windows-laptop loaded this model successfully and the runtime reported phi4-mini-reasoning-q4, but ODS Talk timed out after the model returned no usable agent response; keep it out of agent-required release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-17T00-45-33Z/model-ui/cycle-002/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-17T00-45Z on windows-laptop loaded this model successfully and direct route readiness passed, but ODS Talk did not complete the streamed verification prompt before timeout; keep it out of ODS Talk release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-17T00-45-33Z/model-ui/cycle-002/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "qwen2.5-1.5b-instruct-q4",
      "name": "Qwen 2.5 1.5B Instruct",
      "family": "qwen",
      "gguf_file": "qwen2.5-1.5b-instruct-q4_k_m.gguf",
      "gguf_url": "https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct-GGUF/resolve/91cad51170dc346986eccefdc2dd33a9da36ead9/qwen2.5-1.5b-instruct-q4_k_m.gguf",
      "gguf_sha256": "6a1a2eb6d15622bf3c96857206351ba97e1af16c30d7a74ee38970e434e9407e",
      "size_mb": 1120,
      "vram_required_gb": 3,
      "context_length": 32768,
      "quantization": "Q4_K_M",
      "specialty": "Fast",
      "description": "Qwen's compact 1.5B instruct model with a 32K native context. Good as a very low-VRAM smoke and chat model.",
      "tokens_per_sec_estimate": 160,
      "llm_model_name": "qwen2.5-1.5b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-16T22-53Z on windows-laptop loaded this model successfully and runtime reported qwen2.5-1.5b-instruct-q4, but ODS Talk returned a generic tool-style response instead of the requested verification phrase; keep it out of agent-required release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-16T22-53-30Z-release-affected-product-223c18ff-harness-1faf0d8-windows-laptop/model-ui/cycle-002/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-16T22-53Z on windows-laptop loaded this model successfully and runtime reported qwen2.5-1.5b-instruct-q4, but ODS Talk returned a generic tool-style response instead of the requested verification phrase; keep it out of ODS Talk release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-16T22-53-30Z-release-affected-product-223c18ff-harness-1faf0d8-windows-laptop/model-ui/cycle-002/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "qwen2.5-0.5b-instruct-q4",
      "name": "Qwen 2.5 0.5B Instruct",
      "family": "qwen",
      "gguf_file": "qwen2.5-0.5b-instruct-q4_k_m.gguf",
      "gguf_url": "https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct-GGUF/resolve/9217f5db79a29953eb74d5343926648285ec7e67/qwen2.5-0.5b-instruct-q4_k_m.gguf",
      "gguf_sha256": "74a4da8c9fdbcd15bd1f6d01d621410d31c6fc00986f5eb687824e7b93d7a9db",
      "size_mb": 491,
      "vram_required_gb": 2,
      "context_length": 32768,
      "quantization": "Q4_K_M",
      "specialty": "Fast",
      "description": "Qwen's smallest instruct GGUF with a 32K native context. Useful as a fast, low-risk smoke model on very constrained GPUs.",
      "tokens_per_sec_estimate": 190,
      "llm_model_name": "qwen2.5-0.5b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-16T16-05Z on windows-laptop loaded this model successfully, but ODS Talk did not return the requested verification phrase and asked for clarification instead; keep it out of agent-required release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-16T16-05-53Z-release-affected-product-f22eb7f-harness-c991431-windows-laptop/model-ui/cycle-003/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-16T16-05Z on windows-laptop loaded this model successfully, but ODS Talk did not return the requested verification phrase and asked for clarification instead; keep it out of ODS Talk release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-16T16-05-53Z-release-affected-product-f22eb7f-harness-c991431-windows-laptop/model-ui/cycle-003/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "granite3.3-2b-instruct-q4",
      "name": "Granite 3.3 2B Instruct",
      "family": "granite",
      "gguf_file": "granite-3.3-2b-instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/granite-3.3-2b-instruct-GGUF/resolve/8137ae9f7dd30033d76472c87aa8d41c72cb7c10/granite-3.3-2b-instruct-Q4_K_M.gguf",
      "gguf_sha256": "8ec7f401c35853115d1d94cd7c350f4405f7468589759958eea54bd430c30963",
      "size_bytes": 1545303296,
      "size_mb": 1545,
      "vram_required_gb": 3,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Chat",
      "description": "IBM's compact Granite instruct model with 131K context. Candidate low-VRAM agent model pending fleet validation.",
      "tokens_per_sec_estimate": 145,
      "llm_model_name": "granite-3.3-2b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-16T18-10Z on windows-laptop loaded this model successfully and the runtime reported granite3.3-2b-instruct-q4, but ODS Talk returned a Hermes websocket closed error and then fetch failed during the streamed verification prompt; keep it out of agent-required release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-16T18-10-30Z-release-affected-product-bc4f77d-harness-7630b69-windows-laptop/model-ui/cycle-003/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-16T18-10Z on windows-laptop loaded this model successfully and the runtime reported granite3.3-2b-instruct-q4, but ODS Talk returned a Hermes websocket closed error and then fetch failed during the streamed verification prompt; keep it out of ODS Talk release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-16T18-10-30Z-release-affected-product-bc4f77d-harness-7630b69-windows-laptop/model-ui/cycle-003/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "granite4.0-h-micro-q4",
      "name": "Granite 4.0 H-Micro",
      "family": "granite",
      "gguf_file": "granite-4.0-h-micro-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ibm-granite/granite-4.0-h-micro-GGUF/resolve/dc1dd2585fac18a78001c677d33ef8a7bbb7eb68/granite-4.0-h-micro-Q4_K_M.gguf",
      "gguf_sha256": "bcc78b9b25450101d1ad90d4b9a264e1bac892f534dfb76066f4eec792fdf023",
      "size_bytes": 1942564512,
      "size_mb": 1943,
      "vram_required_gb": 4,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Chat",
      "description": "IBM's compact Granite 4 hybrid instruction model with long-context support. A low-VRAM alternative family for agent and chat validation.",
      "tokens_per_sec_estimate": 115,
      "llm_model_name": "granite-4.0-h-micro",
      "llama_server_image": null,
      "install_recommendation": false
    },
    {
      "id": "granite4.0-h-tiny-q4",
      "name": "Granite 4.0 H-Tiny",
      "family": "granite",
      "gguf_file": "granite-4.0-h-tiny-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ibm-granite/granite-4.0-h-tiny-GGUF/resolve/08d5a8a9741dd5c1a95d2d39e25253226aa1464e/granite-4.0-h-tiny-Q4_K_M.gguf",
      "gguf_sha256": "5a38b08c441ae1adbafb1d2b8a7167e0d48734d83af68b268cefea1eec553dcd",
      "size_bytes": 4230976352,
      "size_mb": 4231,
      "vram_required_gb": 8,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Chat",
      "description": "IBM's larger Granite 4 hybrid instruction model with long-context support. A higher-capacity 8GB-class candidate for agent and chat validation.",
      "tokens_per_sec_estimate": 80,
      "llm_model_name": "granite-4.0-h-tiny",
      "llama_server_image": null,
      "app_compatibility": {
        "pixel_agent": {
          "status": "unsupported_until_revalidated",
          "label": "Pixel revalidation required",
          "reason": "Pixel routed correctly to Granite 4.0 H-Tiny but failed to answer its verification phrase in exact-head release cycles on Tower3 (2026-09-20), Tower2, and the Windows laptop WSL lane (2026-09-21). On Tower2 and Windows it returned delivery instructions instead of the requested phrase, while the exact Granite route was proven and the other required apps passed on the same loaded model; baseline recovery succeeded. Keep this model out of Tower2, Tower3, and Windows laptop all-app release coverage until Pixel behavior is revalidated.",
          "evidence": "ods-public-beta-fleet-green-20260911/release-candidate-01d81ef8-hb71daee-r377/model-ui/cycle-002/tower3/model-ui.json; ods-public-beta-fleet-green-20260911/release-candidate-9381eff2-h6665980-r400/model-ui/cycle-003/tower2/model-ui.json; ods-public-beta-fleet-green-20260911/release-candidate-9381eff2-h6665980-r400/model-ui/cycle-005/windows-laptop-wsl-beta/model-ui.json",
          "recordedAt": "2026-09-21T04:23:13Z",
          "productSha": "9381eff2822390ea0dd44f9471f27c5cffb346e4",
          "harnessSha": "66659802985da37d3289ca33bb815d375b307323",
          "hostScope": ["tower2", "tower3", "windows-laptop"],
          "expiresAt": "2026-10-21T00:00:00Z"
        },
        "opencode": {
          "status": "unsupported_until_revalidated",
          "label": "OpenCode revalidation required",
          "reason": "In the 2026-09-17 Tower1 exact-head release cycle, OpenCode's build agent issued a read tool call against a fabricated path instead of returning its requested verification phrase on Granite 4.0 H-Tiny. The same installed OpenCode path answered on the baseline and neighboring Granite 4.0 models; Pixel, ODS Talk, LiteLLM, Open WebUI, and Perplexica passed on H-Tiny with exact model-route evidence. Keep this model out of Tower1 all-app release coverage until OpenCode behavior is revalidated.",
          "evidence": "ods-public-beta-fleet-green-20260911/release-tower1-pr5579-471bc50-135daf6-r216/model-ui/cycle-002/tower1/model-ui.json",
          "recordedAt": "2026-09-17T08:46:03Z",
          "productSha": "471bc50df3b3a127aa83ac208c7220e19493d8f6",
          "harnessSha": "135daf604399fe757e8364d3721767a22bafe541",
          "hostScope": ["tower1"],
          "expiresAt": "2026-10-17T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "smollm3-3b-q4",
      "name": "SmolLM3 3B",
      "family": "smollm",
      "gguf_file": "SmolLM3-3B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/SmolLM3-3B-GGUF/resolve/a7bc17204c8a326d6bd6e466e076959eddae2025/SmolLM3-3B-Q4_K_M.gguf",
      "gguf_sha256": "4de907d2d388a5508fb7cb443a06effe14cce3518b0a78d3bdd9e74d9edce989",
      "size_bytes": 1915306528,
      "size_mb": 1915,
      "vram_required_gb": 4,
      "context_length": 65536,
      "quantization": "Q4_K_M",
      "specialty": "Chat",
      "description": "Hugging Face's small multilingual long-context instruct model. Candidate low-VRAM agent model pending fleet validation.",
      "tokens_per_sec_estimate": 125,
      "llm_model_name": "smollm3-3b",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "verified",
          "reason": "Fresh exact-commit fleet revalidation on windows-laptop loaded SmolLM3 with its observed 65,536-token runtime context, passed challenge-bound ODS Talk and required deployed LLM-consumer probes, restored the original model and repeated Talk/app probes, then deleted the test model with a clean final inventory.",
          "evidence": "fleet-test/runs/2026-07-24T02-33-53Z-targeted-windows-smollm3-product-58944ba461fb-harness-b838abf5a87b/model-ui/cycle-001/windows-laptop",
          "recordedAt": "2026-07-24T02:40:33Z",
          "productSha": "58944ba461fb87b87b1ce6fa854e32d15aeb8efa",
          "harnessSha": "b838abf5a87b32c3f7ab8eaef43ea97133fb43d9"
        },
        "hermes_talk": {
          "status": "verified",
          "label": "ODS Talk verified",
          "reason": "Fresh exact-commit fleet revalidation on windows-laptop passed challenge-bound ODS Talk on SmolLM3 and again after restoring the original model; all required deployed LLM-consumer probes and cleanup checks also passed.",
          "evidence": "fleet-test/runs/2026-07-24T02-33-53Z-targeted-windows-smollm3-product-58944ba461fb-harness-b838abf5a87b/model-ui/cycle-001/windows-laptop",
          "recordedAt": "2026-07-24T02:40:33Z",
          "productSha": "58944ba461fb87b87b1ce6fa854e32d15aeb8efa",
          "harnessSha": "b838abf5a87b32c3f7ab8eaef43ea97133fb43d9"
        },
        "perplexica": {
          "status": "unsupported_until_revalidated",
          "label": "Perplexica revalidation required",
          "reason": "A fresh exact-commit release run loaded SmolLM3 on all six in-scope hosts. It passed every required app on tower2 and strixy, but Perplexica returned unrelated prose instead of its requested challenge token on strix-halo through Lemonade, spark through Linux llama-server, m5-mbp through Apple llama-server, and windows-laptop through Windows llama-server; those four hosts still passed challenge-bound ODS Talk and every other deployed required LLM consumer, with route evidence naming the exact SmolLM3 runtime. Because the semantic failure spans four hosts and multiple platforms/backends without a stable runtime scope, keep SmolLM3 out of all-app release coverage until Perplexica compatibility is revalidated.",
          "evidence": "fleet-test/runs/2026-07-24T02-46-53Z-release-product-faaf5af99c89-harness-b838abf5a87b-hosts-six-scope-6cycle-switchboard/model-ui/cycle-003/tower2; fleet-test/runs/2026-07-24T02-46-53Z-release-product-faaf5af99c89-harness-b838abf5a87b-hosts-six-scope-6cycle-switchboard/model-ui/cycle-003/strix-halo; fleet-test/runs/2026-07-24T02-46-53Z-release-product-faaf5af99c89-harness-b838abf5a87b-hosts-six-scope-6cycle-switchboard/model-ui/cycle-003/spark; fleet-test/runs/2026-07-24T02-46-53Z-release-product-faaf5af99c89-harness-b838abf5a87b-hosts-six-scope-6cycle-switchboard/model-ui/cycle-003/m5-mbp; fleet-test/runs/2026-07-24T02-46-53Z-release-product-faaf5af99c89-harness-b838abf5a87b-hosts-six-scope-6cycle-switchboard/model-ui/cycle-003/windows-laptop; fleet-test/runs/2026-07-24T02-46-53Z-release-product-faaf5af99c89-harness-b838abf5a87b-hosts-six-scope-6cycle-switchboard/model-ui/cycle-003/strixy",
          "recordedAt": "2026-07-24T04:23:49Z",
          "productSha": "faaf5af99c89e38cc43f18239ea94747f62bf16b",
          "harnessSha": "b838abf5a87b32c3f7ab8eaef43ea97133fb43d9",
          "globalScope": true,
          "expiresAt": "2026-08-24T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "gemma3-4b-it-q4",
      "name": "Gemma 3 4B IT",
      "family": "gemma",
      "gguf_file": "gemma-3-4b-it-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ggml-org/gemma-3-4b-it-GGUF/resolve/d0976223747697cb51e056d85c532013931fe52e/gemma-3-4b-it-Q4_K_M.gguf",
      "gguf_sha256": "882e8d2db44dc554fb0ea5077cb7e4bc49e7342a1f0da57901c0802ea21a0863",
      "size_bytes": 2489757856,
      "size_mb": 2490,
      "vram_required_gb": 4,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Balanced",
      "description": "Google's compact Gemma 3 instruction model with 131K context. A lightweight chat candidate for modest GPUs.",
      "tokens_per_sec_estimate": 110,
      "llm_model_name": "gemma-3-4b-it",
      "llama_server_image": null,
      "app_compatibility": {
        "perplexica": {
          "status": "unsupported_until_revalidated",
          "label": "Perplexica revalidation required",
          "reason": "Independent fleet model-UI runs on windows-laptop, strixy, and tower2 loaded this model and ODS Talk passed, but the Perplexica app probe failed instead of returning the requested verification phrase. The repeated cross-platform semantic failure makes this a model-level incompatibility; keep it out of all-app release coverage until Perplexica compatibility is revalidated.",
          "evidence": "fleet-test/runs/2026-07-19T03-40-01Z-release-product-54b2f703190f-harness-2a6a4e6a55b8-hosts-windows-laptop/model-ui/cycle-003/windows-laptop; fleet-test/runs/2026-07-22T09-42-12Z-release-product-b5bb39f2c0f2-harness-cc7b6148eb2c-hosts-windows-laptop-strixy-m5-mbp/model-ui/cycle-003/strixy; fleet-test/runs/2026-07-23T20-06-24Z-release-product-3675af793193-harness-47eedd6c0b40-hosts-six-scope-6cycle-switchboard/model-ui/cycle-003/tower2",
          "recordedAt": "2026-07-23T21:21:17Z",
          "productSha": "3675af7931935ae0d435a6498c98bc114be209ca",
          "harnessSha": "47eedd6c0b40908272ad6ba324868f4acec2f308",
          "globalScope": true,
          "expiresAt": "2026-08-22T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "granite4.0-h-1b-q4",
      "name": "Granite 4.0 H-1B",
      "family": "granite",
      "gguf_file": "granite-4.0-h-1b-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ibm-granite/granite-4.0-h-1b-GGUF/resolve/c2cb1972f511add21f3bae244990b8ff3a3ffb23/granite-4.0-h-1b-Q4_K_M.gguf",
      "gguf_sha256": "da3d737121a96f3c9a316685212376257a7f167b74380855666dd488d6af3bcb",
      "size_mb": 901,
      "vram_required_gb": 3,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Fast",
      "description": "IBM's small Granite 4 hybrid instruction model with validated long-context support. A very low-VRAM Granite-family candidate for app and agent validation.",
      "tokens_per_sec_estimate": 155,
      "llm_model_name": "granite-4.0-h-1b",
      "llama_server_image": null,
      "app_compatibility": {
        "perplexica": {
          "status": "unsupported_until_revalidated",
          "label": "Perplexica revalidation required",
          "reason": "Perplexica emitted only a partial verification phrase with this model in independent model-UI runs on m5-mbp (2026-07-22) and the Windows WSL beta lane (2026-09-21). In the current WSL run the stream completed and used the exact signed Granite H-1B route, while Pixel, ODS Talk, and the other required consumers passed; baseline recovery succeeded. Keep it out of all-app release coverage on these hosts until Perplexica behavior is revalidated.",
          "evidence": "ods-public-beta-fleet-green-20260911/release-candidate-9381eff2-h6665980-r400/model-ui/cycle-003/windows-laptop-wsl-beta/model-ui.json",
          "recordedAt": "2026-09-21T04:04:00Z",
          "productSha": "9381eff2822390ea0dd44f9471f27c5cffb346e4",
          "harnessSha": "66659802985da37d3289ca33bb815d375b307323",
          "hostScope": ["m5-mbp", "windows-laptop-wsl-beta"],
          "expiresAt": "2026-10-21T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "falcon-h1-1.5b-instruct-q4",
      "name": "Falcon H1 1.5B Instruct",
      "family": "falcon",
      "gguf_file": "Falcon-H1-1.5B-Instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/tiiuae/Falcon-H1-1.5B-Instruct-GGUF/resolve/0d3a6cfe25fb4eeab0153fb8623aac5b69d6bd0a/Falcon-H1-1.5B-Instruct-Q4_K_M.gguf",
      "gguf_sha256": "8b51aa2aa34a0373fd0cd64c02eb91d1bc1da681c09e955ad769d4a9b2d8385f",
      "size_bytes": 944786656,
      "size_mb": 945,
      "vram_required_gb": 3,
      "context_length": 65536,
      "quantization": "Q4_K_M",
      "specialty": "Long Context",
      "description": "TII's compact Falcon H1 hybrid instruct model. A 64K low-VRAM validation candidate with a 256K-capable base model.",
      "tokens_per_sec_estimate": 150,
      "llm_model_name": "Falcon-H1-1.5B-Instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-22T18:45Z on strixy loaded Falcon H1 1.5B through Lemonade and routed ODS Talk through ods/current, but Talk emitted a tool-call payload instead of the requested verification phrase.",
          "evidence": "fleet-test/runs/2026-07-22T18-45-07Z-model-ui-product-31c93e53c4aa-harness-ebdb7f948190-hosts-strixy-talk-runtime-strict-6cycle/cycle-004/strixy",
          "recordedAt": "2026-07-22T19:33:20Z",
          "productSha": "31c93e53c4aa8f9831abdb00516639def985ef34",
          "harnessSha": "ebdb7f9481901a35ca1e5108fe2a1769bab197e8",
          "globalScope": true,
          "expiresAt": "2026-08-22T00:00:00Z"
        },
        "opencode": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-21T11:18Z on tower2 loaded Falcon H1 1.5B successfully and routed OpenCode through ods/current, but the OpenCode UX reply emitted a skill/tool meta-response instead of the requested verification phrase.",
          "evidence": "fleet-test/runs/2026-07-21T11-18-01Z-release-product-c43960145c39-harness-9804e8523a6d-hosts-tower2/model-ui/cycle-004/tower2",
          "recordedAt": "2026-07-21T12:12:38Z",
          "productSha": "c43960145c39c5fa24627bf8b0e98b30c362cc85",
          "harnessSha": "9804e8523a6d11ebeb86eac5f5400cf21fc25b39",
          "hostScope": ["tower2"],
          "expiresAt": "2026-08-21T00:00:00Z"
        },
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Falcon H1 1.5B has failed agent-required browser verification in OpenCode on tower2 and ODS Talk on strixy/Lemonade by emitting tool or meta payloads instead of the requested verification phrase; keep it out of agent-required release coverage until revalidated with a tool-aware Talk/app probe.",
          "evidence": "fleet-test/runs/2026-07-21T11-18-01Z-release-product-c43960145c39-harness-9804e8523a6d-hosts-tower2/model-ui/cycle-004/tower2; fleet-test/runs/2026-07-22T18-45-07Z-model-ui-product-31c93e53c4aa-harness-ebdb7f948190-hosts-strixy-talk-runtime-strict-6cycle/cycle-004/strixy",
          "recordedAt": "2026-07-22T19:33:20Z",
          "productSha": "31c93e53c4aa8f9831abdb00516639def985ef34",
          "harnessSha": "ebdb7f9481901a35ca1e5108fe2a1769bab197e8",
          "globalScope": true,
          "expiresAt": "2026-08-22T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "falcon-h1-3b-instruct-q4",
      "name": "Falcon H1 3B Instruct",
      "family": "falcon",
      "gguf_file": "Falcon-H1-3B-Instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/tiiuae/Falcon-H1-3B-Instruct-GGUF/resolve/18cc9812739f6040f795ddf9d92c9da9a8551572/Falcon-H1-3B-Instruct-Q4_K_M.gguf",
      "gguf_sha256": "1926265eab19b0e47f0c2739e499a18e388373a616d71a218b924d5b95bdf702",
      "size_bytes": 1888612384,
      "size_mb": 1889,
      "vram_required_gb": 4,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Long Context",
      "description": "TII's Falcon H1 3B hybrid instruct model. A long-context low-VRAM validation candidate that preserves the passing Falcon H1 family while replacing Phi-3 Mini in release coverage.",
      "tokens_per_sec_estimate": 115,
      "llm_model_name": "Falcon-H1-3B-Instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-21T12:25Z on tower2 loaded Falcon H1 3B and reached ODS Talk through the browser, but Talk emitted a random_uuid tool-call payload instead of the requested verification phrase.",
          "evidence": "fleet-test/runs/2026-07-21T12-25-39Z-release-product-7df3f218a06c-harness-9804e8523a6d-hosts-tower2/model-ui/cycle-004/tower2",
          "recordedAt": "2026-07-21T14:25:00Z",
          "productSha": "7df3f218a06c386bdb40856c928a0618b8e16b57",
          "harnessSha": "d2a31507fa0652dff4c8cc1892e56ccf1b9102bb",
          "globalScope": true,
          "expiresAt": "2026-08-21T00:00:00Z"
        },
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "The model loaded and routed, but the browser-driven ODS Talk verification failed with a tool-call payload instead of the required answer; keep Falcon H1 3B out of agent-required release coverage until revalidated with a tool-aware Talk probe or updated model metadata.",
          "evidence": "fleet-test/runs/2026-07-21T12-25-39Z-release-product-7df3f218a06c-harness-9804e8523a6d-hosts-tower2/model-ui/cycle-004/tower2",
          "recordedAt": "2026-07-21T14:25:00Z",
          "productSha": "7df3f218a06c386bdb40856c928a0618b8e16b57",
          "harnessSha": "d2a31507fa0652dff4c8cc1892e56ccf1b9102bb",
          "globalScope": true,
          "expiresAt": "2026-08-21T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "nvidia-nemotron3-nano-4b-q4",
      "name": "NVIDIA Nemotron 3 Nano 4B",
      "family": "nemotron",
      "gguf_file": "NVIDIA-Nemotron3-Nano-4B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-4B-GGUF/resolve/ba223d14e45525f7fae81db77ea8cabeb2fc6c25/NVIDIA-Nemotron3-Nano-4B-Q4_K_M.gguf",
      "gguf_sha256": "be5d9a656a51922f24f1f09a759cebb694e1f5d9728bf0ef9f8c972c5a0b5ef2",
      "size_bytes": 2837072864,
      "size_mb": 2838,
      "vram_required_gb": 5,
      "context_length": 65536,
      "max_context_length": 262144,
      "block_count": 42,
      "attention_layer_count": 4,
      "attention_head_count_kv": 8,
      "attention_key_length": 128,
      "attention_value_length": 128,
      "recurrent_state_bytes": 85026816,
      "architecture_source": {
        "repo": "nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16",
        "revision": "dfaf35de3e30f1867dd8dbc38a7fc9fb52d3914f",
        "file": "config.json"
      },
      "selection": {
        "discrete": 45,
        "unified": 50,
        "cpu": 45
      },
      "quantization": "Q4_K_M",
      "specialty": "Agentic Reasoning",
      "description": "NVIDIA's compact Nemotron 3 hybrid reasoning model with long-context and tool-use support. A diverse 5GB-class candidate for local agents, chat, and code.",
      "tokens_per_sec_estimate": 105,
      "llm_model_name": "NVIDIA-Nemotron3-Nano-4B",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "verified",
          "label": "Fleet chat routes verified",
          "reason": "Fresh exact-commit six-host validation loaded NVIDIA Nemotron 3 Nano 4B through the browser switchboard and passed every required deployed OpenAI-compatible LLM-consumer probe with signed route evidence naming the candidate runtime.",
          "evidence": "fleet-test/runs/2026-07-27T02-32-10Z-model-ui-product-651e048b61d9-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T03:04:50Z",
          "productSha": "651e048b61d9a5cc2133f342014b9ceb64800d41",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "hermes_talk": {
          "status": "verified",
          "label": "ODS Talk verified across fleet",
          "reason": "Fresh exact-commit six-host validation passed challenge-bound ODS Talk after loading Nemotron and again after restoring each host's original model.",
          "evidence": "fleet-test/runs/2026-07-27T02-32-10Z-model-ui-product-651e048b61d9-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T03:04:50Z",
          "productSha": "651e048b61d9a5cc2133f342014b9ceb64800d41",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "perplexica": {
          "status": "verified",
          "label": "Perplexica verified across fleet",
          "reason": "Fresh exact-commit six-host validation passed the challenge-bound Perplexica route on Nemotron with signed model-route evidence, then repeated the required consumer suite after clean baseline restoration.",
          "evidence": "fleet-test/runs/2026-07-27T02-32-10Z-model-ui-product-651e048b61d9-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T03:04:50Z",
          "productSha": "651e048b61d9a5cc2133f342014b9ceb64800d41",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "agent_viability": {
          "status": "verified",
          "label": "Agent workflow verified across fleet",
          "reason": "Fresh exact-commit six-host validation passed Nemotron load, challenge-bound ODS Talk, every required deployed LLM consumer, original-model restore, repeated Talk and consumer probes, candidate deletion, and clean final inventory across Linux, macOS, and Windows backends.",
          "evidence": "fleet-test/runs/2026-07-27T02-32-10Z-model-ui-product-651e048b61d9-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T03:04:50Z",
          "productSha": "651e048b61d9a5cc2133f342014b9ceb64800d41",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "pixel_agent": {
          "status": "not_agent_viable",
          "label": "Pixel tool loop failed on Windows",
          "reason": "A real Pixel coding turn created an incorrect TTL/LRU cache, retained two independently reproduced test failures, made no useful repair, and failed to provide a concise honest terminal result. Generic route, Talk, and consumer validation does not establish Pixel capability.",
          "evidence": "docs/pixel/model-capability-qualification.md#nvidia-nemotron-3-nano-4b",
          "recordedAt": "2026-08-30T06:30:00Z",
          "productSha": "df05a732ed7aedac6c527e1f9e7eeeeccfed3a5b",
          "pixelSha": "f1f811d02bffd5a1589eb6feb34323f6dadf7832",
          "harnessSha": "f1f811d02bffd5a1589eb6feb34323f6dadf7832",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-09-30T00:00:00Z"
        }
      },
      "install_recommendation": true
    },
    {
      "id": "granite4.1-3b-q4",
      "name": "Granite 4.1 3B",
      "family": "granite",
      "gguf_file": "granite-4.1-3b-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ibm-granite/granite-4.1-3b-GGUF/resolve/ab4701481089b58a082ef63cc1cee738887293ff/granite-4.1-3b-Q4_K_M.gguf",
      "gguf_sha256": "662b0626cd58f443baea23559b469df6576a81d349649c59413b36a9fb32eb29",
      "size_bytes": 2099501664,
      "size_mb": 2100,
      "vram_required_gb": 4,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Tool Use",
      "description": "IBM's Granite 4.1 3B long-context instruct model quantized for GGUF. A low-VRAM candidate with improved instruction following, chat, and tool-use behavior for app and agent validation.",
      "tokens_per_sec_estimate": 115,
      "llm_model_name": "granite-4.1-3b",
      "llama_server_image": null,
      "app_compatibility": {
        "opencode": {
          "status": "unsupported_until_revalidated",
          "label": "OpenCode revalidation required",
          "reason": "Release-grade fleet evidence has shown intermittent one-character OpenCode answers for Granite 4.1 3B on strix-halo and spark. In the fresh exact-commit Spark failure, ODS Talk, LiteLLM, Open WebUI, and Perplexica passed and routed to the exact llama-server model, but OpenCode stopped after two output tokens and returned only \"O\" instead of its challenge-bound verification phrase. Keep this model out of all-app release coverage on those hosts until the OpenCode path is revalidated.",
          "evidence": "fleet-test/runs/2026-07-23T21-41-59Z-release-product-cf6dffb8e8bb-harness-47eedd6c0b40-hosts-six-scope-6cycle-switchboard/model-ui/cycle-004/strix-halo; fleet-test/runs/2026-07-24T14-11-30Z-release-product-f35e9af8eeab-harness-954deb755b07-hosts-six-scope-6cycle-switchboard/model-ui/cycle-004/spark",
          "recordedAt": "2026-07-24T16:36:20Z",
          "productSha": "f35e9af8eeab7bb154081d1a3076ad9824f58616",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "hostScope": ["strix-halo", "spark"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-22T08:31Z on windows-laptop loaded Granite 4.1 3B and ODS Talk accepted the stream request, but the browser stayed in the Thinking state until the 180s inference timeout while dashboard-api reported repeated llama metrics read timeouts.",
          "evidence": "fleet-test/runs/2026-07-22T08-31-37Z-release-product-52559816f964-harness-ef040aeb79ab-hosts-windows-laptop-strixy-m5-mbp/model-ui/cycle-003/windows-laptop",
          "recordedAt": "2026-07-22T09:18:55Z",
          "productSha": "52559816f964364a63bdbc000722dbbccefec4f9",
          "harnessSha": "ef040aeb79ab1619e8d3d0c7da59baac150e95c6",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-22T00:00:00Z"
        },
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "The model loaded and routed through the browser-managed swap path on windows-laptop, but ODS Talk did not complete a user-facing answer within the release inference window; keep it out of agent-required release coverage on that host until revalidated.",
          "evidence": "fleet-test/runs/2026-07-22T08-31-37Z-release-product-52559816f964-harness-ef040aeb79ab-hosts-windows-laptop-strixy-m5-mbp/model-ui/cycle-003/windows-laptop",
          "recordedAt": "2026-07-22T09:18:55Z",
          "productSha": "52559816f964364a63bdbc000722dbbccefec4f9",
          "harnessSha": "ef040aeb79ab1619e8d3d0c7da59baac150e95c6",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-22T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "granite4.0-1b-q4",
      "name": "Granite 4.0 1B",
      "family": "granite",
      "gguf_file": "granite-4.0-1b-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ibm-granite/granite-4.0-1b-GGUF/resolve/b27c2fe3f211b7f44e80fa620177aea371099aaa/granite-4.0-1b-Q4_K_M.gguf",
      "gguf_sha256": "22ec0f9cc99a90185312de3c882c84e7bd6789bdd050389844380a01a831d7f1",
      "size_mb": 1020,
      "vram_required_gb": 3,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Fast",
      "description": "IBM's compact Granite 4 dense instruction model with validated long-context support. A second small Granite-family candidate for low-VRAM app and agent validation.",
      "tokens_per_sec_estimate": 150,
      "llm_model_name": "granite-4.0-1b",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "verified",
          "reason": "Targeted windows-laptop diagnostic on 2026-07-19 loaded granite4.0-1b-q4 and llama-server returned the requested verification phrase through the OpenAI-compatible chat API, but measured only 0.93 tok/s.",
          "evidence": "fleet-test/runs/2026-07-18T23-55-56Z-release-product-741c5a5360ae-harness-46275424f31d-hosts-windows-laptop/model-ui/cycle-005/windows-laptop plus targeted repro at 2026-07-19T01:15Z"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-18T23-55Z on windows-laptop loaded this model successfully, but ODS Talk timed out after 180s. A targeted direct-chat repro returned the verification phrase at 0.93 tok/s, below the 2.0 tok/s floor for ODS Talk.",
          "evidence": "fleet-test/runs/2026-07-18T23-55-56Z-release-product-741c5a5360ae-harness-46275424f31d-hosts-windows-laptop/model-ui/cycle-005/windows-laptop"
        },
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "The model can answer direct chat, but its measured 0.93 tok/s response rate is below the 2.0 tok/s floor for agent-required workflows and ODS Talk timed out in release coverage.",
          "evidence": "fleet-test/runs/2026-07-18T23-55-56Z-release-product-741c5a5360ae-harness-46275424f31d-hosts-windows-laptop/model-ui/cycle-005/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "granite4.0-h-350m-q4",
      "name": "Granite 4.0 H-350M",
      "family": "granite",
      "gguf_file": "granite-4.0-h-350m-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ibm-granite/granite-4.0-h-350m-GGUF/resolve/a864f823cce6e6048b5752e2816fe7a23987d790/granite-4.0-h-350m-Q4_K_M.gguf",
      "gguf_sha256": "0a8d6a7373602fadfba274a640ba784b86cc6847f1c67f1b0a90fa2ec266b7fb",
      "size_bytes": 222662560,
      "size_mb": 223,
      "vram_required_gb": 2,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Fast",
      "description": "IBM's tiny Granite 4 hybrid instruction model. A smallest-footprint long-context candidate for low-VRAM app and agent validation.",
      "tokens_per_sec_estimate": 190,
      "llm_model_name": "granite-4.0-h-350m",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-19T03:40Z on windows-laptop loaded this model successfully, but ODS Talk returned only \"ODS\" instead of the requested ODSVAL verification phrase; classify it as not agent-viable until revalidated.",
          "evidence": "fleet-test/runs/2026-07-19T03-40-01Z-release-product-54b2f703190f-harness-2a6a4e6a55b8-hosts-windows-laptop/model-ui/cycle-005/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-19T03:40Z on windows-laptop loaded this model successfully, but ODS Talk could not echo the requested ODSVAL verification phrase.",
          "evidence": "fleet-test/runs/2026-07-19T03-40-01Z-release-product-54b2f703190f-harness-2a6a4e6a55b8-hosts-windows-laptop/model-ui/cycle-005/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "granite3.2-2b-instruct-q4",
      "name": "Granite 3.2 2B Instruct",
      "family": "granite",
      "gguf_file": "granite-3.2-2b-instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ibm-research/granite-3.2-2b-instruct-GGUF/resolve/153d944aab8ce9c56ddcedbafc1a682341651f11/granite-3.2-2b-instruct-Q4_K_M.gguf",
      "gguf_sha256": "9bc086149f093169fb8e3e7517cd31752bfd9d70e0e7bb3ab351c0a5386cf8c9",
      "size_bytes": 1545296256,
      "size_mb": 1546,
      "vram_required_gb": 4,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Tool Use",
      "description": "IBM Research's compact Granite 3.2 instruct model with 131K context and tool-use oriented chat template. A low-VRAM long-context candidate for app and agent validation.",
      "tokens_per_sec_estimate": 130,
      "llm_model_name": "granite-3.2-2b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "verified",
          "reason": "Targeted windows-laptop repro on 2026-07-20 loaded granite3.2-2b-instruct-q4 and the OpenAI-compatible chat API returned the requested verification phrase, but measured only 0.73 tok/s.",
          "evidence": "fleet-test/runs/2026-07-20T14-47-35Z-release-product-6b802d2d34a5-harness-d205823ca1d4-hosts-windows-laptop/model-ui/cycle-004/windows-laptop plus targeted repro at 2026-07-20T16:13Z"
        },
        "perplexica": {
          "status": "unsupported_until_revalidated",
          "label": "Perplexica revalidation required",
          "reason": "Fleet runs on Tower2, M5, Tower3, Tower1, and the M4 Mac Mini found Perplexica failures with this model. In the exact 93a2c4bc Mac release run, the signed Granite 3.2 route and every other required consumer passed, but Perplexica produced no headers for five minutes and timed out; exact Qwen 3.5 baseline recovery and candidate deletion then passed. Keep Granite 3.2 2B out of all-app release coverage on these hosts until Perplexica passes a real revalidation.",
          "evidence": "fleet-test/runs/2026-07-23T21-41-59Z-release-product-cf6dffb8e8bb-harness-47eedd6c0b40-hosts-six-scope-6cycle-switchboard/model-ui/cycle-005/tower2; fleet-test/runs/2026-07-23T21-41-59Z-release-product-cf6dffb8e8bb-harness-47eedd6c0b40-hosts-six-scope-6cycle-switchboard/model-ui/cycle-005/m5-mbp; ods-public-beta-fleet-green-20260911/release-four-host-2f985137-abb61b2-r46/model-ui/cycle-006/tower3/model-ui.json; ods-public-beta-fleet-green-20260911/release-tower1-pr5579-afe7408-ff56b22-r208/model-ui/cycle-006/tower1/model-ui.json; ods-public-beta-fleet-green-20260911/release-mac-native-93a2c4bc-h6ab9611-r417/model-ui/cycle-006/mac-mini/model-ui.json",
          "recordedAt": "2026-09-21T15:39:37Z",
          "productSha": "93a2c4bc9b7e8a47335d468b2b4772ead5091680",
          "harnessSha": "6ab9611be19b544e9764c456206d1283f371fbb5",
          "hostScope": ["tower2", "m5-mbp", "tower3", "tower1", "mac-mini"],
          "expiresAt": "2026-10-21T00:00:00Z"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-20T14:47Z on windows-laptop loaded this model successfully, but ODS Talk timed out after 180s. A targeted Talk repro on the same host streamed only a session frame and keepalives for 240s while llama-server was still processing a 19,349-token Hermes prompt.",
          "evidence": "fleet-test/runs/2026-07-20T14-47-35Z-release-product-6b802d2d34a5-harness-d205823ca1d4-hosts-windows-laptop/model-ui/cycle-004/windows-laptop",
          "recordedAt": "2026-07-20T14:47:35Z",
          "productSha": "6b802d2d34a5",
          "harnessSha": "d205823ca1d4",
          "hostScope": ["windows-laptop"],
          "revalidateAgainstSha": "c691f060f93cd0a255760e2749a58c5006edb0c4"
        },
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "The model can answer direct chat, but its measured 0.73 tok/s response rate and 240s ODS Talk timeout under a 19,349-token Hermes prompt make it too slow for agent-required workflows on this host class until revalidated.",
          "evidence": "fleet-test/runs/2026-07-20T14-47-35Z-release-product-6b802d2d34a5-harness-d205823ca1d4-hosts-windows-laptop/model-ui/cycle-004/windows-laptop",
          "recordedAt": "2026-07-20T14:47:35Z",
          "productSha": "6b802d2d34a5",
          "harnessSha": "d205823ca1d4",
          "hostScope": ["windows-laptop"],
          "revalidateAgainstSha": "c691f060f93cd0a255760e2749a58c5006edb0c4"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "granite3.1-2b-instruct-q4",
      "name": "Granite 3.1 2B Instruct",
      "family": "granite",
      "gguf_file": "granite-3.1-2b-instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/bartowski/granite-3.1-2b-instruct-GGUF/resolve/e47b8b46c04cede00f9e19d5a846551b14b2efce/granite-3.1-2b-instruct-Q4_K_M.gguf",
      "gguf_sha256": "774269c82fde2720ea18dcf457fb5bd028fe096139a0735f4ad59c0a270cfc9c",
      "size_bytes": 1545295424,
      "size_mb": 1546,
      "vram_required_gb": 4,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Tool Use",
      "description": "A compact Granite 3.1 instruct quant with 131K context. A same-family fallback candidate for low-VRAM app and agent validation.",
      "tokens_per_sec_estimate": 128,
      "llm_model_name": "granite-3.1-2b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "perplexica": {
          "status": "unsupported_until_revalidated",
          "label": "Perplexica revalidation required",
          "reason": "Fleet model-UI runs on tower2 and strixy loaded this model and ODS Talk, LiteLLM, Open WebUI, and OpenCode passed where deployed, but the Perplexica app probe answered with prose/search-style content instead of the requested verification phrase; keep it out of all-app release coverage until Perplexica compatibility is revalidated.",
          "evidence": "fleet-test/runs/2026-07-21T08-59-51Z-targeted-granite31-allapps-product-73762173ae0c-harness-ecd8157a6324-hosts-tower2/cycle-001/tower2; fleet-test/runs/2026-07-22T21-13-08Z-model-ui-product-39164d8c5bc3-harness-18e2bd60437f-hosts-strixy-switchboard-enabled-6cycle/cycle-006/strixy",
          "recordedAt": "2026-07-22T22:21:04Z",
          "productSha": "39164d8c5bc3306dd4e7ac3e884dcc8574b9930a",
          "harnessSha": "18e2bd60437f4e5ba7b3cd461a37d99ab524ef6a",
          "globalScope": true,
          "expiresAt": "2026-08-22T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "phi3-mini-128k-q4",
      "name": "Phi-3 Mini 128K",
      "family": "phi",
      "gguf_file": "Phi-3-mini-128k-instruct.Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/QuantFactory/Phi-3-mini-128k-instruct-GGUF/resolve/0cc851bb3014ecd709dbe418684c8d66baa35263/Phi-3-mini-128k-instruct.Q4_K_M.gguf",
      "gguf_sha256": "3b27c1a245243b3eadf6db453ddefd419a31e388820824beeba1c60eee17d05e",
      "size_bytes": 2393231712,
      "size_mb": 2393,
      "vram_required_gb": 4,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Long Context",
      "description": "Microsoft Phi-3 Mini 128K instruction model quantized for GGUF. A compact long-context fallback candidate for low-VRAM app and agent validation.",
      "tokens_per_sec_estimate": 105,
      "llm_model_name": "Phi-3-mini-128k-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-21T09:52Z on tower2 loaded Phi-3 Mini 128K successfully, but ODS Talk returned a generic assistant greeting instead of the requested verification phrase; keep it out of ODS Talk release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-21T09-52-42Z-release-product-5d4e805fab79-harness-9804e8523a6d-hosts-tower2/model-ui/cycle-006/tower2",
          "recordedAt": "2026-07-21T11:04:41Z",
          "productSha": "5d4e805fab795bd6c71a805a1c6188852ebb0963",
          "harnessSha": "9804e8523a6d11ebeb86eac5f5400cf21fc25b39",
          "hostScope": ["tower2"],
          "expiresAt": "2026-08-21T00:00:00Z"
        },
        "perplexica": {
          "status": "unsupported_until_revalidated",
          "label": "Perplexica revalidation required",
          "reason": "Strixy browser model-UI coverage loaded Phi-3 Mini 128K and passed ODS Talk, LiteLLM, and Open WebUI, but the Perplexica HTTP 200 reply began with apology prose and omitted the requested verification phrase; keep it out of all-app release coverage until Perplexica compatibility is revalidated.",
          "evidence": "fleet-test/runs/2026-07-23T01-27-32Z-model-ui-product-98d5a50cd5b5-harness-18e2bd60437f-hosts-strixy-switchboard-enabled-6cycle-rerun/cycle-006/strixy",
          "recordedAt": "2026-07-23T02:25:29Z",
          "productSha": "98d5a50cd5b527cb8dbd6117ba2d158538b73a9e",
          "harnessSha": "18e2bd60437f4e5ba7b3cd461a37d99ab524ef6a",
          "globalScope": true,
          "expiresAt": "2026-08-23T00:00:00Z"
        },
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "ODS Talk release probe failed exact verification after successful Phi-3 Mini 128K load; not agent-viable until a Talk revalidation passes.",
          "evidence": "fleet-test/runs/2026-07-21T09-52-42Z-release-product-5d4e805fab79-harness-9804e8523a6d-hosts-tower2/model-ui/cycle-006/tower2",
          "recordedAt": "2026-07-21T11:04:41Z",
          "productSha": "5d4e805fab795bd6c71a805a1c6188852ebb0963",
          "harnessSha": "9804e8523a6d11ebeb86eac5f5400cf21fc25b39",
          "hostScope": ["tower2"],
          "expiresAt": "2026-08-21T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "ministral3-8b-instruct-2512-q4",
      "name": "Ministral 3 8B Instruct 2512",
      "family": "mistral",
      "gguf_file": "Ministral-3-8B-Instruct-2512-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512-GGUF/resolve/0102285ad796bd99af90f58de616092e5630e970/Ministral-3-8B-Instruct-2512-Q4_K_M.gguf",
      "gguf_sha256": "33e7a72cf5e6e2cfc2f2847075acc013d68bba023e35310cef86b5cf8fdca761",
      "size_bytes": 5198911904,
      "size_mb": 5199,
      "vram_required_gb": 14,
      "context_length": 65536,
      "max_context_length": 262144,
      "block_count": 34,
      "attention_layer_count": 34,
      "attention_head_count_kv": 8,
      "attention_key_length": 128,
      "attention_value_length": 128,
      "recurrent_state_bytes": 0,
      "architecture_source": {
        "repo": "mistralai/Ministral-3-8B-Instruct-2512",
        "revision": "5b26027e7b19eeb4b7352e1fed3926375dd2cb4d",
        "file": "config.json"
      },
      "selection": {
        "discrete": 30,
        "unified": 30,
        "cpu": 30
      },
      "quantization": "Q4_K_M",
      "specialty": "Agentic Multilingual",
      "description": "Mistral AI's edge-optimized 8B instruction model with 256K context, native function calling, JSON output, multilingual support, and optional vision through a separate projection artifact.",
      "tokens_per_sec_estimate": 55,
      "llm_model_name": "Ministral-3-8B-Instruct-2512",
      "llama_server_image": null,
      "runtime_profiles": [
        {
          "id": "cpu-16k-agent-memory",
          "label": "CPU 16K agent context with 8G memory budget",
          "backend": "cpu",
          "system_ram_min_gb": 16,
          "context_length": 16384,
          "estimated_required_gb": 7.8,
          "estimate_source": "fleet-validated",
          "fit_label": "16K CPU agent profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "auto",
            "LLAMA_ARG_CACHE_TYPE_K": "f16",
            "LLAMA_ARG_CACHE_TYPE_V": "f16",
            "LLAMA_SERVER_MEMORY_LIMIT": "8G"
          }
        },
        {
          "id": "nvidia-8gb-64k-q4-kv",
          "label": "NVIDIA 8GB 64K context with Q4 KV cache",
          "backend": "nvidia",
          "host_arch": ["amd64"],
          "memory_type": "discrete",
          "vram_min_gb": 7.5,
          "vram_max_gb": 8.5,
          "system_ram_min_gb": 31,
          "context_length": 65536,
          "estimated_required_gb": 7.4,
          "estimate_source": "fleet-validated",
          "fit_label": "64K 8GB Q4 KV profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q4_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q4_0"
          }
        },
        {
          "id": "nvidia-8gb-32k-q8-kv",
          "label": "NVIDIA 8GB 32K context with Q8 KV cache",
          "backend": "nvidia",
          "host_arch": ["amd64"],
          "memory_type": "discrete",
          "vram_min_gb": 7.5,
          "vram_max_gb": 8.5,
          "system_ram_min_gb": 15,
          "context_length": 32768,
          "estimated_required_gb": 6.8,
          "estimate_source": "fleet-validated",
          "fit_label": "32K 8GB Q8 KV constrained-runtime profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0",
            "LLAMA_SERVER_MEMORY_LIMIT": "12G"
          }
        }
      ],
      "app_compatibility": {
        "openai_chat": {
          "status": "verified",
          "label": "Fleet chat routes verified",
          "reason": "Fresh exact-commit six-host validation loaded Ministral 3 8B through the browser switchboard and passed every required deployed OpenAI-compatible LLM-consumer probe with signed route evidence naming the candidate runtime.",
          "evidence": "fleet-test/runs/2026-07-27T06-31-36Z-model-ui-product-7629cd20c0ec-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T06:59:37Z",
          "productSha": "7629cd20c0ec75a274187aea52b8cc9ad6fa2a2a",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "hermes_talk": {
          "status": "verified",
          "label": "ODS Talk verified across fleet",
          "reason": "Fresh exact-commit six-host validation passed challenge-bound ODS Talk after loading Ministral and again after restoring each host's original model.",
          "evidence": "fleet-test/runs/2026-07-27T06-31-36Z-model-ui-product-7629cd20c0ec-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T06:59:37Z",
          "productSha": "7629cd20c0ec75a274187aea52b8cc9ad6fa2a2a",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "perplexica": {
          "status": "verified",
          "label": "Perplexica verified across fleet",
          "reason": "Fresh exact-commit six-host validation passed the challenge-bound Perplexica route on Ministral with signed model-route evidence, then repeated the required consumer suite after clean baseline restoration.",
          "evidence": "fleet-test/runs/2026-07-27T06-31-36Z-model-ui-product-7629cd20c0ec-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T06:59:37Z",
          "productSha": "7629cd20c0ec75a274187aea52b8cc9ad6fa2a2a",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "agent_viability": {
          "status": "verified",
          "label": "Agent workflow verified across fleet",
          "reason": "Fresh exact-commit six-host validation passed Ministral load with the scoped NVIDIA 8GB profile, challenge-bound ODS Talk, every required deployed LLM consumer, original-model restore, repeated Talk and consumer probes, candidate deletion, and clean final inventory across Linux, macOS, and Windows backends.",
          "evidence": "fleet-test/runs/2026-07-27T06-31-36Z-model-ui-product-7629cd20c0ec-harness-19d43e6f9f25-hosts-tower2-strix-halo-spark-m5-mbp-windows-laptop-strixy/cycle-001/{tower2,strix-halo,spark,m5-mbp,windows-laptop,strixy}",
          "recordedAt": "2026-07-27T06:59:37Z",
          "productSha": "7629cd20c0ec75a274187aea52b8cc9ad6fa2a2a",
          "harnessSha": "19d43e6f9f2533e8768ed85b33de9f4ace232129",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop", "strixy"]
        },
        "pixel_agent": {
          "status": "not_agent_viable",
          "label": "Pixel tool loop failed on Windows",
          "reason": "A real Pixel coding turn wrote tests to the wrong directory, ran the wrong verification command, hallucinated a filename correction after an import failure, emitted invalid and ineffective edits, and recursively deleted the requested project tree without owner authorization.",
          "evidence": "docs/pixel/model-capability-qualification.md#ministral-3-8b-instruct-2512",
          "recordedAt": "2026-08-30T06:50:00Z",
          "productSha": "df05a732ed7aedac6c527e1f9e7eeeeccfed3a5b",
          "pixelSha": "f1f811d02bffd5a1589eb6feb34323f6dadf7832",
          "harnessSha": "f1f811d02bffd5a1589eb6feb34323f6dadf7832",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-09-30T00:00:00Z"
        }
      },
      "install_recommendation": true
    },
    {
      "id": "ministral-3b-instruct-q4",
      "name": "Ministral 3B Instruct",
      "family": "mistral",
      "gguf_file": "Ministral-3b-instruct.Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/mradermacher/Ministral-3b-instruct-GGUF/resolve/e31dc5d6827bc82ebf9525f1b6ee9cb33f0ab6a6/Ministral-3b-instruct.Q4_K_M.gguf",
      "gguf_sha256": "eb97f59f4fa2f53b1e585eb36f5388d1ded29a0a167e2d371f2765811853ecb4",
      "size_bytes": 1997338016,
      "size_mb": 1997,
      "vram_required_gb": 4,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Chat",
      "description": "A compact Ministral instruct model with 131K context. A lightweight alternative family for low-VRAM chat validation.",
      "tokens_per_sec_estimate": 120,
      "llm_model_name": "ministral-3b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-17T00-45Z on windows-laptop loaded this model successfully, but ODS Talk returned malformed repetitive output instead of the requested verification phrase; keep it out of agent-required release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-17T00-45-33Z/model-ui/cycle-005/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-17T00-45Z on windows-laptop loaded this model successfully, but ODS Talk returned malformed repetitive output instead of the requested verification phrase; keep it out of ODS Talk release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-17T00-45-33Z/model-ui/cycle-005/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "llama3.2-1b-instruct-q4",
      "name": "Llama 3.2 1B Instruct",
      "family": "llama",
      "gguf_file": "Llama-3.2-1B-Instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Llama-3.2-1B-Instruct-GGUF/resolve/b69aef112e9f895e6f98d7ae0949f72ff09aa401/Llama-3.2-1B-Instruct-Q4_K_M.gguf",
      "gguf_sha256": "3f5a22426976ab26cfe84dba63c1d08391717abb1af893e10f1b2968d862dcc1",
      "size_bytes": 807694368,
      "size_mb": 808,
      "vram_required_gb": 2,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Fast",
      "description": "Meta's smallest Llama 3.2 instruct model with 131K context. Candidate low-VRAM agent model pending fleet validation.",
      "tokens_per_sec_estimate": 170,
      "llm_model_name": "llama-3.2-1b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-16T20-22Z on windows-laptop loaded this model successfully, but ODS Talk returned an invalid tool/function-style payload instead of the requested verification phrase; keep it out of agent-required release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-16T20-22-58Z-release-affected-product-712d2c0e-harness-1faf0d8-windows-laptop/model-ui/cycle-004/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-16T20-22Z on windows-laptop loaded this model successfully, but ODS Talk returned an invalid tool/function-style payload instead of the requested verification phrase; keep it out of ODS Talk release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-16T20-22-58Z-release-affected-product-712d2c0e-harness-1faf0d8-windows-laptop/model-ui/cycle-004/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "llama3.2-3b-instruct-q4",
      "name": "Llama 3.2 3B Instruct",
      "family": "llama",
      "gguf_file": "llama-3.2-3b-instruct-q4_k_m.gguf",
      "gguf_url": "https://huggingface.co/hugging-quants/Llama-3.2-3B-Instruct-Q4_K_M-GGUF/resolve/eb72f2a08dd2b9edd07ffacfe5aa56938b7939b0/llama-3.2-3b-instruct-q4_k_m.gguf",
      "gguf_sha256": "c55a83bfb6396799337853ca69918a0b9bbb2917621078c34570bc17d20fd7a1",
      "size_bytes": 2019373920,
      "size_mb": 2020,
      "vram_required_gb": 4,
      "context_length": 128000,
      "quantization": "Q4_K_M",
      "specialty": "Fast",
      "description": "Meta's small instruction model. Good for fast chat and compatibility checks on low-VRAM machines.",
      "tokens_per_sec_estimate": 120,
      "llm_model_name": "llama-3.2-3b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-14T01-31Z on strix-halo loaded this model successfully but ODS Talk refused the verification-code prompt instead of answering; keep it out of agent-required release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-14T01-31-17Z-release-affected-product-4ca225b8-harness-09f8f6f-tower2-strixhalo"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-14T01-31Z on strix-halo loaded this model successfully but ODS Talk refused the verification-code prompt instead of answering; keep it out of agent-required release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-14T01-31-17Z-release-affected-product-4ca225b8-harness-09f8f6f-tower2-strixhalo"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "qwen2.5-3b-instruct-q4",
      "name": "Qwen 2.5 3B Instruct",
      "family": "qwen",
      "gguf_file": "qwen2.5-3b-instruct-q4_k_m.gguf",
      "gguf_url": "https://huggingface.co/Qwen/Qwen2.5-3B-Instruct-GGUF/resolve/7dabda4d13d513e3e842b20f0d435c732f172cbe/qwen2.5-3b-instruct-q4_k_m.gguf",
      "gguf_sha256": "626b4a6678b86442240e33df819e00132d3ba7dddfe1cdc4fbb18e0a9615c62d",
      "size_bytes": 2104932768,
      "size_mb": 2105,
      "vram_required_gb": 4,
      "context_length": 32768,
      "quantization": "Q4_K_M",
      "specialty": "Chat",
      "description": "Compact general-purpose Qwen instruction model with 32K runtime context in the current ODS llama.cpp path. Useful for direct chat, but below the 64K floor for agent-required release coverage.",
      "tokens_per_sec_estimate": 112,
      "llm_model_name": "qwen2.5-3b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "verified",
          "reason": "Fleet model-UI run 2026-07-18T21-03Z on windows-laptop loaded this model and completed the required OpenAI-compatible app probes, but runtime metadata reported 32,768 context instead of the previous cataloged 128K value.",
          "evidence": "fleet-test/runs/2026-07-18T21-03-13Z-model-ui-product-137d335d3231-harness-46275424f31d-hosts-windows-laptop/cycle-004/windows-laptop"
        },
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "The same fleet run reported only 32,768 runtime context for this GGUF on the Windows llama.cpp path, below the 64K ODS agent-release floor; keep it out of agent-required release coverage until the context source of truth is resolved.",
          "evidence": "fleet-test/runs/2026-07-18T21-03-13Z-model-ui-product-137d335d3231-harness-46275424f31d-hosts-windows-laptop/cycle-004/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Runtime context is below the 64K ODS Talk release floor on the Windows llama.cpp path, even though short app probes completed.",
          "evidence": "fleet-test/runs/2026-07-18T21-03-13Z-model-ui-product-137d335d3231-harness-46275424f31d-hosts-windows-laptop/cycle-004/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "qwen3-4b-q4",
      "name": "Qwen3 4B",
      "family": "qwen",
      "gguf_file": "Qwen3-4B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/Qwen/Qwen3-4B-GGUF/resolve/bc640142c66e1fdd12af0bd68f40445458f3869b/Qwen3-4B-Q4_K_M.gguf",
      "gguf_sha256": "7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5",
      "size_mb": 2500,
      "vram_required_gb": 5,
      "context_length": 40960,
      "quantization": "Q4_K_M",
      "specialty": "Chat",
      "description": "Qwen's compact 4B chat model with a 40K runtime context cap in the current ODS llama.cpp path. It remains available for manual testing, but is blocked from agent-required release coverage until long-context and app behavior are revalidated.",
      "tokens_per_sec_estimate": 1,
      "llm_model_name": "qwen3-4b",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-18T21-03Z on windows-laptop loaded and routed this model correctly, but llama.cpp capped context at 40,960 and ODS Talk timed out after 180s while runtime decode was about 0.5 tok/s; keep it out of agent-required release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-18T21-03-13Z-model-ui-product-137d335d3231-harness-46275424f31d-hosts-windows-laptop/cycle-005/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "ODS Talk timed out after the model loaded and routed correctly on windows-laptop, with runtime context capped at 40,960 and decode around 0.5 tok/s.",
          "evidence": "fleet-test/runs/2026-07-18T21-03-13Z-model-ui-product-137d335d3231-harness-46275424f31d-hosts-windows-laptop/cycle-005/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "qwen3-4b-instruct-2507-q4",
      "name": "Qwen3 4B Instruct 2507",
      "family": "qwen",
      "gguf_file": "Qwen3-4B-Instruct-2507-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF/resolve/a06e946bb6b655725eafa393f4a9745d460374c9/Qwen3-4B-Instruct-2507-Q4_K_M.gguf",
      "gguf_sha256": "3605803b982cb64aead44f6c1b2ae36e3acdb41d8e46c8a94c6533bc4c67e597",
      "size_bytes": 2497281120,
      "size_mb": 2497,
      "vram_required_gb": 5,
      "context_length": 262144,
      "quantization": "Q4_K_M",
      "specialty": "Chat",
      "description": "Unsloth's GGUF quant of Qwen3 4B Instruct 2507 with 262K context metadata. A low-VRAM long-context release candidate for app and Talk validation.",
      "tokens_per_sec_estimate": 105,
      "llm_model_name": "qwen3-4b-instruct-2507",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "verified",
          "label": "Direct chat verified on Windows",
          "reason": "Fresh exact-commit Windows fleet validation loaded the profiled model through the UI and passed challenge-bound LiteLLM and Open WebUI routes with exact model-route evidence.",
          "evidence": "fleet-test/runs/2026-07-24T07-55-00Z-windows-profile-candidates-route-validation/cycle-001/windows-laptop",
          "recordedAt": "2026-07-24T08:03:57Z",
          "productSha": "449cf84d866d8bdedd8046d3c58faab6c07b5f03",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "hermes_talk": {
          "status": "verified",
          "label": "ODS Talk verified on Windows",
          "reason": "Fresh exact-commit Windows fleet validation passed challenge-bound ODS Talk with the NVIDIA 8GB 64K Q4-KV profile, then restored the original model and passed Talk again.",
          "evidence": "fleet-test/runs/2026-07-24T07-55-00Z-windows-profile-candidates-route-validation/cycle-001/windows-laptop",
          "recordedAt": "2026-07-24T08:03:57Z",
          "productSha": "449cf84d866d8bdedd8046d3c58faab6c07b5f03",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "perplexica": {
          "status": "verified",
          "label": "Perplexica verified on Windows",
          "reason": "Fresh exact-commit Windows fleet validation passed the challenge-bound Perplexica app route with exact Qwen3 4B Instruct route evidence before clean baseline restore and candidate deletion.",
          "evidence": "fleet-test/runs/2026-07-24T07-55-00Z-windows-profile-candidates-route-validation/cycle-001/windows-laptop",
          "recordedAt": "2026-07-24T08:03:57Z",
          "productSha": "449cf84d866d8bdedd8046d3c58faab6c07b5f03",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "agent_viability": {
          "status": "verified",
          "label": "Agent workflow verified on Windows",
          "reason": "Fresh exact-commit Windows fleet validation loaded Qwen3 4B Instruct with the NVIDIA 8GB 64K Q4-KV profile and passed ODS Talk plus every required deployed LLM-consumer probe in test and restored-baseline phases, with clean candidate deletion and final inventory.",
          "evidence": "fleet-test/runs/2026-07-24T07-55-00Z-windows-profile-candidates-route-validation/cycle-001/windows-laptop",
          "recordedAt": "2026-07-24T08:03:57Z",
          "productSha": "449cf84d866d8bdedd8046d3c58faab6c07b5f03",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "pixel_agent": {
          "status": "not_agent_viable",
          "label": "Pixel repair loop failed on Windows",
          "reason": "A real Pixel coding turn at 24K context ignored the deterministic fake-clock contract, wrote defective cache code and tests, independently retained two failures out of seven, and then made ineffective and duplicate repairs until the bounded verification guard terminated the continuation honestly. Direct tool calls and generic route, Talk, or consumer evidence do not establish Pixel capability.",
          "evidence": "docs/pixel/model-capability-qualification.md#qwen3-4b-instruct-2507",
          "recordedAt": "2026-08-30T08:10:12Z",
          "productSha": "df05a732ed7aedac6c527e1f9e7eeeeccfed3a5b",
          "pixelSha": "f1f811d02bffd5a1589eb6feb34323f6dadf7832",
          "harnessSha": "f1f811d02bffd5a1589eb6feb34323f6dadf7832",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-09-30T00:00:00Z"
        }
      },
      "runtime_profiles": [
        {
          "id": "nvidia-8gb-64k-q4-kv",
          "label": "NVIDIA 8GB 64K context with Q4 KV cache",
          "backend": "nvidia",
          "host_arch": ["amd64"],
          "memory_type": "discrete",
          "vram_min_gb": 7.5,
          "vram_max_gb": 8.5,
          "system_ram_min_gb": 31,
          "context_length": 65536,
          "estimated_required_gb": 7.2,
          "estimate_source": "fleet-validated",
          "fit_label": "64K 8GB Q4 KV profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q4_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q4_0"
          }
        }
      ],
      "install_recommendation": false
    },
    {
      "id": "qwen3-4b-128k-q4",
      "name": "Qwen3 4B 128K",
      "family": "qwen",
      "gguf_file": "Qwen3-4B-128K-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3-4B-128K-GGUF/resolve/a21842b9582c3888b30568b8e84afe533275f646/Qwen3-4B-128K-Q4_K_M.gguf",
      "gguf_sha256": "f145a1bd60fec420ca4d9b7645ebcdf657e301463bc4dd3af4a8c0b548b5eb1a",
      "size_bytes": 2497281504,
      "size_mb": 2497,
      "vram_required_gb": 5,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Chat",
      "description": "Unsloth's 128K GGUF build of Qwen3 4B in Q4_K_M quantization. A low-VRAM long-context candidate kept separate from the upstream 40K-context Qwen3 4B entry.",
      "tokens_per_sec_estimate": 105,
      "llm_model_name": "qwen3-4b-128k",
      "llama_server_image": null,
      "app_compatibility": {
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "label": "ODS Talk revalidation required",
          "reason": "Fresh exact-commit release runs loaded Qwen3 4B 128K through Linux llama-server on tower2, Apple llama-server on m5-mbp, and Windows llama-server on windows-laptop. Tower2 loaded and routed the exact model in 3.9 seconds, but Talk replied that it generated the challenge phrase without actually returning it. M5 returned only a generated .MEDIA audio marker and omitted the challenge text, while Windows did not finish within the enforced 180-second inference limit. Keep this model out of Talk/all-app release coverage globally until challenge-bound response behavior is revalidated.",
          "evidence": "fleet-test/runs/2026-07-24T02-46-53Z-release-product-faaf5af99c89-harness-b838abf5a87b-hosts-six-scope-6cycle-switchboard/model-ui/cycle-006/m5-mbp; fleet-test/runs/2026-07-24T02-46-53Z-release-product-faaf5af99c89-harness-b838abf5a87b-hosts-six-scope-6cycle-switchboard/model-ui/cycle-005/windows-laptop; fleet-test/runs/2026-07-24T10-53-17Z-release-product-68d41e774eb1-harness-954deb755b07-hosts-six-scope-6cycle-switchboard/model-ui/cycle-006/tower2",
          "recordedAt": "2026-07-24T12:06:17Z",
          "productSha": "68d41e774eb13be334aa8cd3904ecbadaf69f4bf",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "globalScope": true,
          "expiresAt": "2026-08-24T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "qwen3-1.7b-q4",
      "name": "Qwen3 1.7B",
      "family": "qwen",
      "gguf_file": "Qwen3-1.7B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ggml-org/Qwen3-1.7B-GGUF/resolve/daeb8e2d528a760970442092f6bf1e55c3b659eb/Qwen3-1.7B-Q4_K_M.gguf",
      "gguf_sha256": "d2387ca2dbfee2ffabce7120d3770dadca0b293052bc2f0e138fdc940d9bc7b5",
      "size_mb": 1282,
      "vram_required_gb": 3,
      "context_length": 40960,
      "quantization": "Q4_K_M",
      "specialty": "Fast",
      "description": "Qwen3's compact 1.7B chat model in Q4_K_M quantization. The upstream config advertises 40K default context without YaRN, below the 64K ODS agent-release floor.",
      "tokens_per_sec_estimate": 145,
      "llm_model_name": "qwen3-1.7b",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Upstream Qwen3 1.7B config reports max_position_embeddings=40960 and no rope_scaling; keep it out of agent-required release coverage until ODS has a product-visible YaRN/runtime policy for this model family.",
          "evidence": "https://huggingface.co/Qwen/Qwen3-1.7B/blob/main/config.json"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Default runtime context is below the 64K ODS Talk release floor until a YaRN/runtime policy is implemented and fleet-validated.",
          "evidence": "https://huggingface.co/Qwen/Qwen3-1.7B/blob/main/config.json"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "qwen2.5-coder-1.5b-128k-q4",
      "name": "Qwen 2.5 Coder 1.5B 128K",
      "family": "qwen",
      "gguf_file": "Qwen2.5-Coder-1.5B-Instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen2.5-Coder-1.5B-Instruct-128K-GGUF/resolve/15e05dbeacf9e5dc519ee1ac64308465cee5db77/Qwen2.5-Coder-1.5B-Instruct-Q4_K_M.gguf",
      "gguf_sha256": "0fbff4d39395fab063c51377ba522928af2574b1f998d66012c1caed7b8f91d6",
      "size_bytes": 986048224,
      "size_mb": 986,
      "vram_required_gb": 3,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Code",
      "description": "Small code-specialized Qwen model with a 128K GGUF context. A low-VRAM OpenCode-oriented release candidate to replace models that route correctly but fail agent UX probes.",
      "tokens_per_sec_estimate": 145,
      "llm_model_name": "qwen2.5-coder-1.5b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-21T12:25Z on tower2 loaded Qwen 2.5 Coder 1.5B 128K, but ODS Talk answered with generic assistant prose instead of the requested verification phrase.",
          "evidence": "fleet-test/runs/2026-07-21T12-25-39Z-release-product-7df3f218a06c-harness-9804e8523a6d-hosts-tower2/model-ui/cycle-006/tower2",
          "recordedAt": "2026-07-21T14:25:00Z",
          "productSha": "7df3f218a06c386bdb40856c928a0618b8e16b57",
          "harnessSha": "d2a31507fa0652dff4c8cc1892e56ccf1b9102bb",
          "hostScope": ["tower2"],
          "expiresAt": "2026-08-21T00:00:00Z"
        },
        "opencode": {
          "status": "unsupported_until_revalidated",
          "label": "OpenCode revalidation required",
          "reason": "Fresh exact-commit release coverage on strix-halo and spark routed OpenCode to Qwen 2.5 Coder 1.5B 128K after challenge-bound ODS Talk, LiteLLM, and Open WebUI passed. The strix-halo run emitted a webfetch tool payload with a mangled verification token; the spark run replied that no changes were made and omitted the requested token while Perplexica also passed. Keep this model out of OpenCode release coverage on those hosts until revalidated.",
          "evidence": "fleet-test/runs/2026-07-24T12-22-22Z-release-product-9ad366f47050-harness-954deb755b07-hosts-six-scope-6cycle-switchboard/model-ui/cycle-006/strix-halo; fleet-test/runs/2026-07-24T20-41-29Z-release-product-8318f6cf37d8-harness-7b83ca3a780f-hosts-six-scope-6cycle-switchboard/model-ui/cycle-006/spark",
          "recordedAt": "2026-07-24T22:11:24Z",
          "productSha": "8318f6cf37d878ea5d31dab0f557678f1e8ca3ef",
          "harnessSha": "7b83ca3a780f502f6f12c90596256fe1cb8235af",
          "hostScope": ["strix-halo", "spark"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "perplexica": {
          "status": "unsupported_until_revalidated",
          "label": "Perplexica revalidation required",
          "reason": "Fresh exact-route browser coverage on strix-halo, m5-mbp, and windows-laptop loaded Qwen 2.5 Coder 1.5B 128K and passed challenge-bound ODS Talk, LiteLLM, and Open WebUI. Perplexica returned unrelated research output and omitted its requested verification phrase on all three hosts. Keep this model out of Perplexica release coverage on those hosts until revalidated.",
          "evidence": "fleet-test/runs/2026-07-24T12-22-22Z-release-product-9ad366f47050-harness-954deb755b07-hosts-six-scope-6cycle-switchboard/model-ui/cycle-006/strix-halo; fleet-test/runs/2026-07-24T20-41-29Z-release-product-8318f6cf37d8-harness-7b83ca3a780f-hosts-six-scope-6cycle-switchboard/model-ui/cycle-005/m5-mbp; fleet-test/runs/2026-07-24T18-04-guardrail-diagnostic-windows-qwen25-coder-product-f35e9af8eeab-harness-7c6cd3af853b/cycle-001/windows-laptop",
          "recordedAt": "2026-07-24T22:22:05Z",
          "productSha": "8318f6cf37d878ea5d31dab0f557678f1e8ca3ef",
          "harnessSha": "7b83ca3a780f502f6f12c90596256fe1cb8235af",
          "hostScope": ["strix-halo", "m5-mbp", "windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "The model failed agent-required browser verification on tower2 in ODS Talk, on strix-halo in OpenCode and Perplexica, on spark in OpenCode, and on m5-mbp and windows-laptop in Perplexica after exact-model routing. Keep it out of agent-required release coverage on those hosts until the affected app paths are revalidated.",
          "evidence": "fleet-test/runs/2026-07-21T12-25-39Z-release-product-7df3f218a06c-harness-9804e8523a6d-hosts-tower2/model-ui/cycle-006/tower2; fleet-test/runs/2026-07-24T12-22-22Z-release-product-9ad366f47050-harness-954deb755b07-hosts-six-scope-6cycle-switchboard/model-ui/cycle-006/strix-halo; fleet-test/runs/2026-07-24T20-41-29Z-release-product-8318f6cf37d8-harness-7b83ca3a780f-hosts-six-scope-6cycle-switchboard/model-ui/cycle-006/spark; fleet-test/runs/2026-07-24T20-41-29Z-release-product-8318f6cf37d8-harness-7b83ca3a780f-hosts-six-scope-6cycle-switchboard/model-ui/cycle-005/m5-mbp; fleet-test/runs/2026-07-24T18-04-guardrail-diagnostic-windows-qwen25-coder-product-f35e9af8eeab-harness-7c6cd3af853b/cycle-001/windows-laptop",
          "recordedAt": "2026-07-24T22:22:05Z",
          "productSha": "8318f6cf37d878ea5d31dab0f557678f1e8ca3ef",
          "harnessSha": "7b83ca3a780f502f6f12c90596256fe1cb8235af",
          "hostScope": ["tower2", "strix-halo", "spark", "m5-mbp", "windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "qwen2.5-coder-3b-128k-q4",
      "name": "Qwen 2.5 Coder 3B 128K",
      "family": "qwen",
      "gguf_file": "Qwen2.5-Coder-3B-Instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen2.5-Coder-3B-Instruct-128K-GGUF/resolve/5326551926d06f7f9cab53c9b9b552e3bedfe8ba/Qwen2.5-Coder-3B-Instruct-Q4_K_M.gguf",
      "gguf_sha256": "a77ce2f29052145ce425f2e3bb6f7359be66ae8eb2a73ab032bebe2f6bf22e9d",
      "size_bytes": 1929902688,
      "size_mb": 1930,
      "vram_required_gb": 4,
      "context_length": 128000,
      "quantization": "Q4_K_M",
      "specialty": "Code",
      "description": "Compact code-specialized Qwen model with 128K context. Strong candidate for developer workflows on modest GPUs.",
      "tokens_per_sec_estimate": 115,
      "llm_model_name": "qwen2.5-coder-3b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "verified",
          "label": "Agent workflow verified on Windows",
          "reason": "Targeted windows-laptop browser revalidation with the literal-response persona guardrail loaded the exact Qwen 2.5 Coder 3B 128K route and passed challenge-bound ODS Talk plus every deployed LLM consumer, baseline restore, deletion, inventory, and environment cleanup.",
          "evidence": "fleet-test/runs/2026-07-24T18-23-guardrail-fullapp-windows-qwen25-coder-3b-product-f35e9af8eeab-harness-7c6cd3af853b/cycle-001/windows-laptop",
          "recordedAt": "2026-07-24T18:30:00Z",
          "hostScope": ["windows-laptop"]
        },
        "pixel_agent": {
          "status": "not_agent_viable",
          "label": "Pixel protocol failed on Windows",
          "reason": "A real Pixel coding turn made no tool calls, created no files, and returned malformed mock example protocol text instead of operating the requested workspace.",
          "evidence": "docs/pixel/model-capability-qualification.md#qwen-25-coder-3b-128k",
          "recordedAt": "2026-08-30T06:25:00Z",
          "productSha": "df05a732ed7aedac6c527e1f9e7eeeeccfed3a5b",
          "pixelSha": "f1f811d02bffd5a1589eb6feb34323f6dadf7832",
          "harnessSha": "f1f811d02bffd5a1589eb6feb34323f6dadf7832",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-09-30T00:00:00Z"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "label": "ODS Talk revalidation required on Tower2 and M5",
          "reason": "Fresh exact-commit release coverage on tower2 loaded Qwen 2.5 Coder 3B 128K through the browser switchboard, but ODS Talk acknowledged that the verification phrase was correct without returning the challenge token. Final15 fresh exact-commit release coverage on m5-mbp loaded the same model in cycle 5, but ODS Talk returned reasoning/probe prose instead of the exact ODSVAL-8DB490CA32-3FC971DE20 verification phrase; cleanup succeeded, the original baseline was restored, the test model was deleted, and cycle 6 passed. Keep this model out of Tower2 and M5 ODS Talk release coverage until its literal-response behavior is revalidated. The prior windows-laptop full-app pass remains recorded by the host-scoped agent-viability verdict.",
          "evidence": "fleet-test/runs/2026-07-24T23-54-27Z-release-product-b77641dd064f-harness-3930eeed3597-hosts-six-scope-6cycle-switchboard/model-ui/cycle-006/tower2; fleet-test/runs/2026-07-25T18-19-44Z-release-product-d5ef0b7ddfc8-harness-8b9f1a38027e-hosts-six-scope-6cycle-switchboard/model-ui/cycle-005/m5-mbp",
          "recordedAt": "2026-07-25T18:19:44Z",
          "productSha": "d5ef0b7ddfc8207600ec60813e5574dc516c5918",
          "harnessSha": "8b9f1a38027ebafaad074eb7964ff65e7aa1e731",
          "hostScope": ["tower2", "m5-mbp"],
          "expiresAt": "2026-08-25T00:00:00Z"
        },
        "opencode": {
          "status": "unsupported_until_revalidated",
          "label": "OpenCode revalidation required on Strix Halo and Spark",
          "reason": "Fresh exact-commit release coverage on strix-halo and spark loaded Qwen 2.5 Coder 3B 128K through the browser switchboard and passed challenge-bound ODS Talk, LiteLLM, Open WebUI, and Perplexica, but OpenCode did not return the exact unpredictable phrase. Strix Halo returned only the verification prefix; Spark changed the required 'ODS verify' prefix to 'ODES verify'. Keep this model out of Strix Halo and Spark all-app release coverage until its literal-response behavior is revalidated. Prior Strix Halo evidence: fleet-test/runs/2026-07-25T02-56-13Z-release-product-89c6e46fde88-harness-78aeb37d418e-hosts-six-scope-6cycle-switchboard/model-ui/cycle-006/strix-halo.",
          "evidence": "fleet-test/runs/2026-07-25T07-36-57Z-release-product-ac90d567a6ea-harness-bfe040f32941-hosts-six-scope-6cycle-switchboard/model-ui/cycle-006/spark",
          "recordedAt": "2026-07-25T09:13:26Z",
          "productSha": "ac90d567a6ea721cedc2817f98edca16dd60dd0a",
          "harnessSha": "bfe040f32941fa3cf82f709a02ab65dcf2e28c7a",
          "hostScope": ["strix-halo", "spark"],
          "expiresAt": "2026-08-25T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "qwen2.5-7b-instruct-q4",
      "name": "Qwen 2.5 7B Instruct",
      "family": "qwen",
      "gguf_file": "qwen2.5-7b-instruct-q4_k_m-00001-of-00002.gguf",
      "gguf_url": "",
      "gguf_sha256": "",
      "gguf_parts": [
        {
          "file": "qwen2.5-7b-instruct-q4_k_m-00001-of-00002.gguf",
          "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GGUF/resolve/bb5d59e06d9551d752d08b292a50eb208b07ab1f/qwen2.5-7b-instruct-q4_k_m-00001-of-00002.gguf",
          "sha256": "dfce12e3862a5283ccfb88221b48480e58745165de856439950d0f22590580db",
          "size_bytes": 3993201344
        },
        {
          "file": "qwen2.5-7b-instruct-q4_k_m-00002-of-00002.gguf",
          "url": "https://huggingface.co/Qwen/Qwen2.5-7B-Instruct-GGUF/resolve/bb5d59e06d9551d752d08b292a50eb208b07ab1f/qwen2.5-7b-instruct-q4_k_m-00002-of-00002.gguf",
          "sha256": "539cf93f78e887edea1c04e2d7d8cdaca9d01dae9c9025bcb8accbe29df3d72a",
          "size_bytes": 689872288
        }
      ],
      "size_bytes": 4683073632,
      "size_mb": 4683,
      "vram_required_gb": 8,
      "context_length": 128000,
      "quantization": "Q4_K_M",
      "specialty": "Balanced",
      "description": "General-purpose Qwen instruction model with 128K context. A mid-size option for 8GB-class systems.",
      "tokens_per_sec_estimate": 80,
      "llm_model_name": "qwen2.5-7b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-16T11-44Z on windows-laptop loaded this model successfully, but ODS Talk timed out and the 8GB Windows host stopped responding on dashboard-api, llama-server, and SSH before cleanup; keep it out of agent-required release coverage until revalidated on 8GB-class Windows hardware.",
          "evidence": "fleet-test/runs/2026-07-16T11-44-05Z-release-affected-product-d11ef611-harness-a7745b8-windows-pair/model-ui/cycle-005/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-16T11-44Z on windows-laptop loaded this model successfully, but ODS Talk timed out and the 8GB Windows host stopped responding before cleanup; keep it out of ODS Talk release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-16T11-44-05Z-release-affected-product-d11ef611-harness-a7745b8-windows-pair/model-ui/cycle-005/windows-laptop"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "llama3.1-8b-instruct-q4",
      "name": "Llama 3.1 8B Instruct",
      "family": "llama",
      "gguf_file": "Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/bartowski/Meta-Llama-3.1-8B-Instruct-GGUF/resolve/bf5b95e96dac0462e2a09145ec66cae9a3f12067/Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf",
      "gguf_sha256": "7b064f5842bf9532c91456deda288a1b672397a54fa729aa665952863033557c",
      "size_bytes": 4920739232,
      "size_mb": 4921,
      "vram_required_gb": 8,
      "context_length": 128000,
      "quantization": "Q4_K_M",
      "specialty": "General",
      "description": "Meta's 8B long-context instruct model. Good general chat quality while still fitting mainstream GPUs.",
      "tokens_per_sec_estimate": 75,
      "llm_model_name": "llama-3.1-8b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-14T08-31Z on strix-halo loaded this model successfully on the correct Lemonade route, but ODS Talk answered the neutral repeat prompt with a session/tool meta-response instead of the requested phrase; keep it out of agent-required release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-14T08-31-52Z-release-affected-product-2b74ffc7-harness-dbf0068-strixhalo/model-ui/cycle-005/strix-halo"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-14T08-31Z on strix-halo loaded this model successfully on the correct Lemonade route, but ODS Talk answered the neutral repeat prompt with a session/tool meta-response instead of the requested phrase; keep it out of ODS Talk release coverage until revalidated.",
          "evidence": "fleet-test/runs/2026-07-14T08-31-52Z-release-affected-product-2b74ffc7-harness-dbf0068-strixhalo/model-ui/cycle-005/strix-halo"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "granite3.3-8b-instruct-q4",
      "name": "Granite 3.3 8B Instruct",
      "family": "granite",
      "gguf_file": "granite-3.3-8b-instruct-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/ibm-granite/granite-3.3-8b-instruct-GGUF/resolve/e40e9dd739c7be00fa965c16ce167088190ce114/granite-3.3-8b-instruct-Q4_K_M.gguf",
      "gguf_sha256": "77bcee066a76dcdd10d0d123c87e32c8ec2c74e31b6ffd87ebee49c9ac215dca",
      "size_bytes": 4942873344,
      "size_mb": 4943,
      "vram_required_gb": 8,
      "context_length": 128000,
      "quantization": "Q4_K_M",
      "specialty": "Enterprise",
      "description": "IBM's efficient 8B instruct model with 128K context. Useful for business and tool-heavy tasks.",
      "tokens_per_sec_estimate": 72,
      "llm_model_name": "granite-3.3-8b-instruct",
      "llama_server_image": null,
      "app_compatibility": {
        "agent_viability": {
          "status": "not_agent_viable",
          "reason": "Fleet model-UI run 2026-07-19T03:40Z on windows-laptop loaded this model through the NVIDIA 8GB 64K profile, but ODS Talk timed out after 180s; keep it out of release-grade agent/app coverage until revalidated or profiled differently.",
          "evidence": "fleet-test/runs/2026-07-19T03-40-01Z-release-product-54b2f703190f-harness-2a6a4e6a55b8-hosts-windows-laptop/model-ui/cycle-006/windows-laptop"
        },
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "reason": "Fleet model-UI run 2026-07-19T03:40Z on windows-laptop loaded this model through the NVIDIA 8GB 64K profile, but ODS Talk timed out after 180s.",
          "evidence": "fleet-test/runs/2026-07-19T03-40-01Z-release-product-54b2f703190f-harness-2a6a4e6a55b8-hosts-windows-laptop/model-ui/cycle-006/windows-laptop"
        }
      },
      "runtime_profiles": [
        {
          "id": "nvidia-8gb-64k",
          "label": "NVIDIA 8GB 64K context",
          "backend": "nvidia",
          "memory_type": "discrete",
          "vram_min_gb": 7.5,
          "vram_max_gb": 8.5,
          "context_length": 65536,
          "estimated_required_gb": 6.8,
          "fit_label": "64K 8GB profile"
        }
      ],
      "install_recommendation": false
    },
    {
      "id": "mistral-nemo-12b-instruct-q4",
      "name": "Mistral Nemo 12B Instruct",
      "family": "mistral",
      "gguf_file": "Mistral-Nemo-Instruct-2407.Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/QuantFactory/Mistral-Nemo-Instruct-2407-GGUF/resolve/6d01a4babf90e5b1f274c8f37867a6b0494cdf5a/Mistral-Nemo-Instruct-2407.Q4_K_M.gguf",
      "gguf_sha256": "cbe6b246b096352563579207b499aae8da5e9b590b63e72a89dd17db590840d7",
      "size_bytes": 7477204512,
      "size_mb": 7477,
      "vram_required_gb": 12,
      "context_length": 128000,
      "quantization": "Q4_K_M",
      "specialty": "Quality",
      "description": "Mistral's 12B long-context model. A higher-quality option for 12GB+ GPUs and Apple unified memory systems.",
      "tokens_per_sec_estimate": 55,
      "llm_model_name": "mistral-nemo-instruct-2407",
      "llama_server_image": null,
      "app_compatibility": {
        "hermes_talk": {
          "status": "unsupported_until_revalidated",
          "label": "ODS Talk unavailable on Apple llama-server",
          "reason": "Fleet model-UI run 2026-07-22T06:12Z on m5-mbp loaded Mistral Nemo 12B successfully through the browser UI, but ODS Talk failed because llama.cpp rejected the Hermes request with the Mistral chat-template role alternation Jinja exception. The same run passed Mistral on strix-halo through Lemonade, so this verdict is scoped to Apple llama-server until revalidated.",
          "evidence": "fleet-test/runs/2026-07-22T06-12-53Z-release-product-cfe6addc5239-harness-8bdceb20b959-hosts-m5-mbp-strix-halo-windows-laptop-strixy/model-ui/cycle-006/m5-mbp",
          "recordedAt": "2026-07-22T07:47:48Z",
          "productSha": "cfe6addc523925adba4e6ab797645e83741759b0",
          "harnessSha": "8bdceb20b9594515dae65f7f66df734ec152e1e4",
          "hostScope": ["m5-mbp"],
          "gpuBackendScope": ["apple"],
          "llmBackendScope": ["llama-server"],
          "expiresAt": "2026-08-22T00:00:00Z"
        }
      },
      "install_recommendation": false
    },
    {
      "id": "qwen3.5-4b-q4",
      "name": "Qwen 3.5 4B",
      "family": "qwen",
      "gguf_file": "Qwen3.5-4B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/e87f176479d0855a907a41277aca2f8ee7a09523/Qwen3.5-4B-Q4_K_M.gguf",
      "gguf_sha256": "00fe7986ff5f6b463e62455821146049db6f9313603938a70800d1fb69ef11a4",
      "size_bytes": 2740937888,
      "size_mb": 2741,
      "vram_required_gb": 5,
      "context_length": 65536,
      "max_context_length": 262144,
      "block_count": 32,
      "attention_layer_count": 8,
      "attention_head_count_kv": 4,
      "attention_key_length": 256,
      "attention_value_length": 256,
      "recurrent_state_bytes": 52690944,
      "architecture_source": {
        "repo": "Qwen/Qwen3.5-4B",
        "revision": "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a",
        "file": "config.json"
      },
      "selection": {
        "discrete": 50,
        "unified": 40,
        "cpu": 50
      },
      "quantization": "Q4_K_M",
      "specialty": "Balanced",
      "description": "Unsloth's GGUF quant of Qwen3.5 4B with native 262K context metadata. A low-VRAM long-context release candidate for app and Talk validation.",
      "tokens_per_sec_estimate": 100,
      "llm_model_name": "qwen3.5-4b",
      "llama_server_image": null,
      "app_compatibility": {
        "openai_chat": {
          "status": "verified",
          "label": "Direct chat verified on Windows",
          "reason": "Fresh exact-commit Windows fleet validation loaded the profiled model through the UI and passed challenge-bound LiteLLM and Open WebUI routes with exact model-route evidence.",
          "evidence": "fleet-test/runs/2026-07-24T07-55-00Z-windows-profile-candidates-route-validation/cycle-002/windows-laptop",
          "recordedAt": "2026-07-24T08:03:57Z",
          "productSha": "449cf84d866d8bdedd8046d3c58faab6c07b5f03",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "hermes_talk": {
          "status": "verified",
          "label": "ODS Talk verified on Windows",
          "reason": "Fresh exact-commit Windows fleet validation passed challenge-bound ODS Talk with the NVIDIA 8GB 64K Q4-KV profile, then restored the original model and passed Talk again.",
          "evidence": "fleet-test/runs/2026-07-24T07-55-00Z-windows-profile-candidates-route-validation/cycle-002/windows-laptop",
          "recordedAt": "2026-07-24T08:03:57Z",
          "productSha": "449cf84d866d8bdedd8046d3c58faab6c07b5f03",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "perplexica": {
          "status": "verified",
          "label": "Perplexica verified on Windows",
          "reason": "Fresh exact-commit Windows fleet validation passed the challenge-bound Perplexica app route with exact Qwen3.5 4B route evidence before clean baseline restore and candidate deletion.",
          "evidence": "fleet-test/runs/2026-07-24T07-55-00Z-windows-profile-candidates-route-validation/cycle-002/windows-laptop",
          "recordedAt": "2026-07-24T08:03:57Z",
          "productSha": "449cf84d866d8bdedd8046d3c58faab6c07b5f03",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "agent_viability": {
          "status": "verified",
          "label": "Agent workflow verified on Windows",
          "reason": "Fresh exact-commit Windows fleet validation loaded Qwen3.5 4B with the NVIDIA 8GB 64K Q4-KV profile and passed ODS Talk plus every required deployed LLM-consumer probe in test and restored-baseline phases, with clean candidate deletion and final inventory.",
          "evidence": "fleet-test/runs/2026-07-24T07-55-00Z-windows-profile-candidates-route-validation/cycle-002/windows-laptop",
          "recordedAt": "2026-07-24T08:03:57Z",
          "productSha": "449cf84d866d8bdedd8046d3c58faab6c07b5f03",
          "harnessSha": "954deb755b0730719512ac3675a748474180e01c",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-08-24T00:00:00Z"
        },
        "pixel_agent": {
          "status": "not_agent_viable",
          "label": "Pixel repair loop failed on Windows",
          "reason": "A real Pixel coding turn reached a unittest NameError, made one ineffective duplicate edit, exhausted its bounded tool loop, and ended without rerunning verification or reporting the remaining failure honestly. Independent replay retained the error and found additional contract defects.",
          "evidence": "docs/pixel/model-capability-qualification.md#qwen-35-4b",
          "recordedAt": "2026-08-30T06:20:00Z",
          "productSha": "df05a732ed7aedac6c527e1f9e7eeeeccfed3a5b",
          "pixelSha": "f1f811d02bffd5a1589eb6feb34323f6dadf7832",
          "harnessSha": "f1f811d02bffd5a1589eb6feb34323f6dadf7832",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-09-30T00:00:00Z"
        }
      },
      "runtime_profiles": [
        {
          "id": "nvidia-8gb-64k-q4-kv",
          "label": "NVIDIA 8GB 64K context with Q4 KV cache",
          "backend": "nvidia",
          "host_arch": ["amd64"],
          "memory_type": "discrete",
          "vram_min_gb": 7.5,
          "vram_max_gb": 8.5,
          "system_ram_min_gb": 8,
          "context_length": 65536,
          "estimated_required_gb": 7.2,
          "estimate_source": "fleet-validated",
          "fit_label": "64K 8GB Q4 KV profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q4_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q4_0",
            "LLAMA_ARG_CTX_CHECKPOINTS": "8",
            "LLAMA_ARG_CACHE_RAM": "2048"
          }
        },
        {
          "id": "cpu-64k-q8-kv",
          "label": "CPU 64K context with Q8 KV cache and capped checkpoints",
          "backend": "cpu",
          "system_ram_min_gb": 12,
          "system_ram_max_gb": 22,
          "context_length": 65536,
          "fit_label": "64K CPU Q8 KV profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0",
            "LLAMA_ARG_CTX_CHECKPOINTS": "4",
            "LLAMA_ARG_CACHE_RAM": "1024",
            "LLAMA_SERVER_MEMORY_LIMIT": "6G"
          }
        },
        {
          "id": "nvidia-6gb-64k-q8-kv",
          "label": "NVIDIA 4.5-7.5 GB 64K context with Q8 KV cache",
          "backend": "nvidia",
          "host_arch": ["amd64"],
          "memory_type": "discrete",
          "vram_min_gb": 4.5,
          "vram_max_gb": 7.49,
          "system_ram_min_gb": 8,
          "context_length": 65536,
          "fit_label": "64K 6GB Q8 KV profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0"
          }
        }
      ]
    },
    {
      "id": "gemma4-e2b-q4",
      "name": "Gemma 4 E2B",
      "family": "gemma4",
      "gguf_file": "gemma-4-E2B-it-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/gemma-4-E2B-it-GGUF/resolve/0314792d7f1f7e229411f620751375812bb9faf2/gemma-4-E2B-it-Q4_K_M.gguf",
      "gguf_sha256": "740185b21d22ceb83a11c3aa62ad5842ef32c70f6096d756bbee85a1e4ec34b8",
      "size_bytes": 3106738272,
      "size_mb": 2963,
      "vram_required_gb": 5,
      "context_length": 65536,
      "max_context_length": 131072,
      "block_count": 35,
      "attention_layer_count": 3,
      "attention_head_count_kv": 1,
      "attention_key_length": 512,
      "attention_value_length": 512,
      "sliding_window": 512,
      "sliding_window_layer_count": 12,
      "sliding_window_head_count_kv": 1,
      "sliding_window_key_length": 256,
      "sliding_window_value_length": 256,
      "recurrent_state_bytes": 0,
      "architecture_source": {
        "repo": "google/gemma-4-E2B-it",
        "revision": "3e22461f65e89153144f8adb70e3b8c2cc9845a7",
        "file": "config.json",
        "note": "20 of 35 layers reuse earlier layers' KV (num_kv_shared_layers), so 15 layers hold a cache. Sliding-window layers hold only the window (llama.cpp b9014 caches n_swa + n_ubatch cells for them), so KV grows only on the full-attention layers."
      },
      "selection": {
        "discrete": 40,
        "unified": 40,
        "cpu": 40
      },
      "quantization": "Q4_K_M",
      "specialty": "Fast",
      "description": "Google's efficient small model. Good for quick tasks on budget hardware.",
      "tokens_per_sec_estimate": 110,
      "llm_model_name": "gemma-4-e2b-it",
      "llama_server_image": "ghcr.io/ggml-org/llama.cpp:server-cuda-b9014@sha256:fcf285820892e7ce3218379634e3590826fc697e8b6745b9392072462e355c4f"
    },
    {
      "id": "deepseek-r1-7b-q4",
      "name": "DeepSeek R1 7B",
      "family": "deepseek",
      "gguf_file": "DeepSeek-R1-Distill-Qwen-7B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/DeepSeek-R1-Distill-Qwen-7B-GGUF/resolve/097680e4eed7a83b3df6b0bb5e5134099cadf1b0/DeepSeek-R1-Distill-Qwen-7B-Q4_K_M.gguf",
      "gguf_sha256": "78272d8d32084548bd450394a560eb2d70de8232ab96a725769b1f9171235c1c",
      "size_bytes": 4683073248,
      "size_mb": 4680,
      "vram_required_gb": 9,
      "context_length": 65536,
      "max_context_length": 131072,
      "block_count": 28,
      "attention_head_count_kv": 4,
      "attention_key_length": 128,
      "attention_value_length": 128,
      "recurrent_state_bytes": 0,
      "architecture_source": {
        "repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
        "revision": "916b56a44061fd5cd7d6a8fb632557ed4f724f60",
        "file": "config.json"
      },
      "install_recommendation": false,
      "install_recommendation_reason": "No tool training; it only won as the RAM-gate fallback on 8 GB cards. Kept for manual activation.",
      "quantization": "Q4_K_M",
      "specialty": "Reasoning",
      "description": "DeepSeek R1 distilled into Qwen 7B. Chain-of-thought reasoning on a budget GPU.",
      "tokens_per_sec_estimate": 80,
      "llm_model_name": "deepseek-r1-distill-qwen-7b",
      "llama_server_image": null
    },
    {
      "id": "gemma4-e4b-q4",
      "name": "Gemma 4 E4B",
      "family": "gemma4",
      "gguf_file": "gemma-4-E4B-it-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/gemma-4-E4B-it-GGUF/resolve/bfc15c382204943c3a8fff0c750b94ae2364d7a3/gemma-4-E4B-it-Q4_K_M.gguf",
      "gguf_sha256": "85a896a047553e842f25297ee5b031d64ff30147d9c4af17b1e4b394cd1fab87",
      "size_bytes": 4977171584,
      "size_mb": 4747,
      "vram_required_gb": 8,
      "context_length": 65536,
      "max_context_length": 131072,
      "block_count": 42,
      "attention_layer_count": 4,
      "attention_head_count_kv": 2,
      "attention_key_length": 512,
      "attention_value_length": 512,
      "sliding_window": 512,
      "sliding_window_layer_count": 20,
      "sliding_window_head_count_kv": 2,
      "sliding_window_key_length": 256,
      "sliding_window_value_length": 256,
      "recurrent_state_bytes": 0,
      "architecture_source": {
        "repo": "google/gemma-4-E4B-it",
        "revision": "ee0ef6023621cff504d758262d4e04895a5af4a2",
        "file": "config.json",
        "note": "18 of 42 layers reuse earlier layers' KV (num_kv_shared_layers), so 24 layers hold a cache. Sliding-window layers hold only the window (llama.cpp b9014 caches n_swa + n_ubatch cells for them), so KV grows only on the full-attention layers."
      },
      "selection": {
        "discrete": 50,
        "unified": 50,
        "cpu": 50
      },
      "quantization": "Q4_K_M",
      "specialty": "General",
      "description": "Google's mid-range model with strong multilingual and reasoning capabilities.",
      "tokens_per_sec_estimate": 85,
      "llm_model_name": "gemma-4-e4b-it",
      "llama_server_image": "ghcr.io/ggml-org/llama.cpp:server-cuda-b9014@sha256:fcf285820892e7ce3218379634e3590826fc697e8b6745b9392072462e355c4f"
    },
    {
      "id": "qwen3.5-9b-q4",
      "name": "Qwen 3.5 9B",
      "family": "qwen",
      "gguf_file": "Qwen3.5-9B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3.5-9B-GGUF/resolve/3885219b6810b007914f3a7950a8d1b469d598a5/Qwen3.5-9B-Q4_K_M.gguf",
      "gguf_sha256": "03b74727a860a56338e042c4420bb3f04b2fec5734175f4cb9fa853daf52b7e8",
      "size_bytes": 5680522464,
      "size_mb": 5760,
      "vram_required_gb": 8,
      "context_length": 65536,
      "max_context_length": 262144,
      "block_count": 32,
      "attention_layer_count": 8,
      "attention_head_count_kv": 4,
      "attention_key_length": 256,
      "attention_value_length": 256,
      "recurrent_state_bytes": 52690944,
      "architecture_source": {
        "repo": "Qwen/Qwen3.5-9B",
        "revision": "c202236235762e1c871ad0ccb60c8ee5ba337b9a",
        "file": "config.json"
      },
      "selection": {
        "discrete": 60,
        "unified": 60,
        "cpu": 60
      },
      "quantization": "Q4_K_M",
      "specialty": "General",
      "description": "Excellent general-purpose model. Default for most systems. Great at coding, reasoning, and multilingual tasks.",
      "tokens_per_sec_estimate": 90,
      "llm_model_name": "qwen3.5-9b",
      "llama_server_image": null,
      "app_compatibility": {
        "open_webui": {
          "status": "verified",
          "label": "Open WebUI revalidated on Strixy",
          "reason": "The exact e4fd1cae Strixy cycle loaded Qwen 3.5 9B through native Lemonade at 64K context and passed the challenge-bound Open WebUI reply with signed route evidence naming Qwen3.5-9B-GGUF. ODS Talk, Pixel retained and non-retained inference, LiteLLM, OpenCode, and Perplexica also passed before exact Qwen 3.6 baseline restoration and clean inventory verification.",
          "evidence": "ods-public-beta-fleet-green-20260911/release-strixy-e4fd1cae-h3043027-r401/model-ui/cycle-006/strixy-wsl-beta/model-ui.json",
          "recordedAt": "2026-09-21T07:48:20Z",
          "productSha": "e4fd1cae6232af0a9d7438317002b55f94bea2bc",
          "harnessSha": "30430271ba0b97c83056272a4f4877cb65dee122",
          "hostScope": ["strixy"],
          "expiresAt": "2026-10-21T00:00:00Z"
        },
        "pixel_agent": {
          "status": "verified",
          "label": "Pixel verified at 64K context on Windows laptop",
          "reason": "Fresh Pixel sessions on the auto-paired 64K profile created and debugged a new CLI, then repaired two seeded implementation defects while preserving fixed tests. The load-bearing repair ran the exact 12-second unittest command as a background process, polled an initial four-failure terminal result, made only two focused implementation edits, polled the rerun to 9/9 and exit zero, and passed an exact functional marker plus independent replay. This is capability evidence for the named host, artifact, profile, ODS source, and Pixel source; the earlier failed 24K, 32K, and 64K probes remain preserved as historical evidence rather than an access restriction.",
          "evidence": "docs/pixel/model-capability-qualification.md#qwen-35-9b-revalidation-2026-09-02",
          "recordedAt": "2026-09-02T23:04:24Z",
          "productSha": "d0808d08645841ffcbb3cf3919a9c81fe485937b",
          "pixelSha": "d99923246e5ea22c0f1c8c8fc7b0927ac8b523fe",
          "harnessSha": "d99923246e5ea22c0f1c8c8fc7b0927ac8b523fe",
          "hostScope": ["windows-laptop"],
          "expiresAt": "2026-10-02T23:04:24Z"
        }
      },
      "runtime_profiles": [
        {
          "id": "nvidia-8gb-64k-q8-kv",
          "label": "NVIDIA 8GB 64K context with Q8 KV cache",
          "backend": "nvidia",
          "host_arch": ["amd64"],
          "memory_type": "discrete",
          "vram_min_gb": 7.5,
          "vram_max_gb": 8.5,
          "system_ram_min_gb": 15,
          "context_length": 65536,
          "estimated_required_gb": 7.2,
          "estimate_source": "fleet-validated",
          "fit_label": "64K 8GB Q8 KV interactive-agent profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0",
            "LLAMA_SERVER_MEMORY_LIMIT": "12G"
          }
        },
        {
          "id": "nvidia-8gb-32k-q8-kv",
          "label": "NVIDIA 8GB 32K context with Q8 KV cache",
          "backend": "nvidia",
          "host_arch": ["amd64"],
          "memory_type": "discrete",
          "vram_min_gb": 7.5,
          "vram_max_gb": 8.5,
          "system_ram_min_gb": 15,
          "context_length": 32768,
          "estimated_required_gb": 6.8,
          "estimate_source": "fleet-validated",
          "fit_label": "32K 8GB Q8 KV interactive-agent profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0",
            "LLAMA_SERVER_MEMORY_LIMIT": "12G"
          }
        },
        {
          "id": "cpu-64k-q8-kv",
          "label": "CPU 64K context with Q8 KV cache and capped checkpoints",
          "backend": "cpu",
          "system_ram_min_gb": 23,
          "context_length": 65536,
          "fit_label": "64K CPU Q8 KV profile",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0",
            "LLAMA_ARG_CTX_CHECKPOINTS": "4",
            "LLAMA_ARG_CACHE_RAM": "1024",
            "LLAMA_SERVER_MEMORY_LIMIT": "9G"
          }
        },
        {
          "id": "amd-8gb-64k-q8-kv",
          "label": "AMD 8GB 64K context with Q8 KV cache",
          "backend": "amd",
          "memory_type": "discrete",
          "vram_min_gb": 7.5,
          "vram_max_gb": 8.5,
          "context_length": 65536,
          "fit_label": "64K 8GB Q8 KV profile",
          "notes": "These llama.cpp cache settings apply while this model is active; switching to a model without a profile restores the defaults.",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0"
          }
        }
      ]
    },
    {
      "id": "phi4-q4",
      "name": "Phi-4 14B",
      "family": "phi",
      "gguf_file": "phi-4-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/bartowski/phi-4-GGUF/resolve/19cd65f97c2f1712a81c506611d3f9c94b16a1e1/phi-4-Q4_K_M.gguf",
      "gguf_sha256": "009aba717c09d4a35890c7d35eb59d54e1dba884c7c526e7197d9c13ab5911d9",
      "size_bytes": 9053114816,
      "size_mb": 9050,
      "vram_required_gb": 13,
      "context_length": 16384,
      "max_context_length": 16384,
      "block_count": 40,
      "attention_head_count_kv": 10,
      "attention_key_length": 128,
      "attention_value_length": 128,
      "recurrent_state_bytes": 0,
      "architecture_source": {
        "repo": "microsoft/phi-4",
        "revision": "2db69c1c3e91a05d2c64a3185acfbaf36f744e25",
        "file": "config.json"
      },
      "app_compatibility": {
        "hermes_talk": {
          "status": "unsupported",
          "label": "Context too small for ODS Talk",
          "reason": "config.json max_position_embeddings is 16,384, below Hermes's 65,536 minimum context.",
          "userNote": "ODS Talk needs at least 64K of context; this model supports only 16K.",
          "evidence": "https://huggingface.co/microsoft/phi-4/blob/2db69c1c3e91a05d2c64a3185acfbaf36f744e25/config.json"
        }
      },
      "install_recommendation": false,
      "install_recommendation_reason": "Native context is 16,384 and its chat template has no tool calling; at the 64K Hermes floor it needs about 21.4 GiB. Kept for manual activation.",
      "quantization": "Q4_K_M",
      "specialty": "General",
      "description": "Microsoft's MIT-licensed 14B model. Strong at STEM, coding, and structured reasoning.",
      "tokens_per_sec_estimate": 35,
      "llm_model_name": "phi-4",
      "llama_server_image": null
    },
    {
      "id": "deepseek-r1-14b-q4",
      "name": "DeepSeek R1 14B",
      "family": "deepseek",
      "gguf_file": "DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF/resolve/7b05b58b41f623e66fc74cd27b35475267b2f3e3/DeepSeek-R1-Distill-Qwen-14B-Q4_K_M.gguf",
      "gguf_sha256": "67a7933cf2ad596a393c8e13b30bc4da2d50b283e250b78554aed18817eca31c",
      "size_bytes": 8988109984,
      "size_mb": 8990,
      "vram_required_gb": 21,
      "context_length": 65536,
      "max_context_length": 131072,
      "block_count": 48,
      "attention_head_count_kv": 8,
      "attention_key_length": 128,
      "attention_value_length": 128,
      "recurrent_state_bytes": 0,
      "architecture_source": {
        "repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B",
        "revision": "1df8507178afcc1bef68cd8c393f61a886323761",
        "file": "config.json"
      },
      "install_recommendation": false,
      "install_recommendation_reason": "January 2025 reasoning distill with no tool training; next in line on 12-16 GB once phi-4 is gone. Kept for manual activation.",
      "quantization": "Q4_K_M",
      "specialty": "Reasoning",
      "description": "DeepSeek R1 distilled into Qwen 14B. Strong chain-of-thought reasoning for mid-range GPUs.",
      "tokens_per_sec_estimate": 50,
      "llm_model_name": "deepseek-r1-distill-qwen-14b",
      "llama_server_image": null
    },
    {
      "id": "qwen3.8-27b-iq4-xs",
      "name": "Qwen 3.8 27B",
      "family": "qwen",
      "gguf_file": "Qwen3.8-27B-UD-IQ4_XS.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3.8-27B-GGUF/resolve/4ca720788d1e01f1bff70c033e0d0028fd02e502/Qwen3.8-27B-UD-IQ4_XS.gguf",
      "gguf_sha256": "40fac4050e940397dbf13087afd50f4734a11805bf9d65ef8ddd7483470e6199",
      "size_bytes": 14252845984,
      "size_mb": 14253,
      "vram_required_gb": 18,
      "context_length": 65536,
      "max_context_length": 262144,
      "block_count": 64,
      "attention_layer_count": 16,
      "attention_head_count_kv": 4,
      "attention_key_length": 256,
      "attention_value_length": 256,
      "recurrent_state_bytes": 156893184,
      "architecture_source": {
        "repo": "Qwen/Qwen3.8-27B",
        "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
        "file": "config.json"
      },
      "quantization": "UD-IQ4_XS",
      "specialty": "Reasoning and coding",
      "description": "Dense 27B reasoning model with embedded MTP. Smaller GPUs can use partial offload and a reduced context profile. MTP needs a compatible llama.cpp runtime and measured memory/throughput validation; it is not enabled automatically.",
      "llm_model_name": "qwen3.8-27b",
      "llama_server_image": null,
      "runtime_profiles": [
        {
          "id": "nvidia-24gb-64k-b11429",
          "label": "NVIDIA 24GB 64K context with Q8 KV cache",
          "backend": "nvidia",
          "memory_type": "discrete",
          "host_arch": ["amd64"],
          "vram_min_gb": 24,
          "context_length": 65536,
          "fit_label": "64K NVIDIA Q8 KV profile",
          "llama_server_image": "ghcr.io/ggml-org/llama.cpp:server-cuda-b11429@sha256:883915ea20a4b350e98f7284098f4d400c21e9b366e248d397cd42d90602a84c",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0",
            "LLAMA_ARG_CTX_CHECKPOINTS": "4",
            "LLAMA_ARG_CACHE_RAM": "1024",
            "LLAMA_SERVER_MEMORY_LIMIT": "26G"
          }
        },
        {
          "id": "cpu-32gb-64k-b11429",
          "label": "CPU 32GB RAM 64K context with Q8 KV cache",
          "backend": "cpu",
          "host_arch": ["amd64"],
          "system_ram_min_gb": 32,
          "context_length": 65536,
          "fit_label": "64K CPU Q8 KV profile",
          "llama_server_image": "ghcr.io/ggml-org/llama.cpp:server-b11429@sha256:2f8ebc2dde83c4bd7e5c6d61efa57965e0d5a79d29802c246c4593740d9c04c1",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0",
            "LLAMA_ARG_CTX_CHECKPOINTS": "4",
            "LLAMA_ARG_CACHE_RAM": "1024",
            "LLAMA_SERVER_MEMORY_LIMIT": "26G"
          }
        }
      ],
      "install_recommendation": false,
      "default_runtime_compatibility": {
        "status": "incompatible",
        "runtime": "llama.cpp b9014",
        "reason": "The historical llama.cpp b9014 loader failed Qwen3.8-27B UD-Q4_K_XL with missing tensor blk.64.ssm_conv1d.weight. That verdict does not qualify a different build or quantization. The exact catalog IQ4_XS artifact loaded and completed a basic CPU request on b11429 on 2026-10-21; its explicit backend-specific profiles retain separate context and agent acceptance requirements. Backends without a matching image profile retain this refusal.",
        "evidence": "ODS-Fleet-Qualification findings/llama-cpp-version-policy.md, section 3 (tower2 A/B bench, 2026-09-25)",
        "userNote": "This model has no compatible runtime profile for this machine. Choose another model for now."
      },
      "source_repo": "unsloth/Qwen3.8-27B-GGUF",
      "source_revision": "4ca720788d1e01f1bff70c033e0d0028fd02e502",
      "source_url": "https://huggingface.co/unsloth/Qwen3.8-27B-GGUF",
      "license": "apache-2.0",
      "mtp": {
        "mode": "embedded",
        "source_url": "https://unsloth.ai/docs/models/mtp",
        "runtime_requirement": "llama.cpp advertising draft-mtp and --spec-draft-n-max",
        "recommendation": "benchmark-required"
      }
    },
    {
      "id": "qwen3.5-27b-q4",
      "name": "Qwen 3.5 27B",
      "family": "qwen",
      "gguf_file": "Qwen3.5-27B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3.5-27B-GGUF/resolve/3221f178a6b842d04f1fb42f1c413534adcc0a6a/Qwen3.5-27B-Q4_K_M.gguf",
      "gguf_sha256": "84b5f7f112156d63836a01a69dc3f11a6ba63b10a23b8ca7a7efaf52d5a2d806",
      "size_bytes": 16740812704,
      "size_mb": 16700,
      "vram_required_gb": 21,
      "context_length": 65536,
      "max_context_length": 262144,
      "block_count": 64,
      "attention_layer_count": 16,
      "attention_head_count_kv": 4,
      "attention_key_length": 256,
      "attention_value_length": 256,
      "recurrent_state_bytes": 156893184,
      "architecture_source": {
        "repo": "Qwen/Qwen3.5-27B",
        "revision": "fc05daec18b0a78c049392ed2e771dde82bdf654",
        "file": "config.json"
      },
      "selection": {
        "discrete": 80,
        "unified": 0,
        "cpu": 0
      },
      "quantization": "Q4_K_M",
      "specialty": "Quality",
      "description": "Dense 27B powerhouse. Top-tier quality for 24GB GPUs. Excels at complex reasoning and coding.",
      "tokens_per_sec_estimate": 25,
      "llm_model_name": "qwen3.5-27b",
      "llama_server_image": null,
      "runtime_profiles": [
        {
          "id": "nvidia-20gb-64k-q8-kv",
          "label": "NVIDIA 20GB 64K context with Q8 KV cache",
          "backend": "nvidia",
          "memory_type": "discrete",
          "vram_min_gb": 19.5,
          "vram_max_gb": 23.49,
          "context_length": 65536,
          "fit_label": "64K 20GB Q8 KV profile",
          "notes": "Measured on an RTX 5090 at llama.cpp b9014: 18,742 MiB with Q8 KV at 64K, 2% slower decode than F16 KV.",
          "env": {
            "LLAMA_PARALLEL": "1",
            "LLAMA_ARG_FLASH_ATTN": "on",
            "LLAMA_ARG_CACHE_TYPE_K": "q8_0",
            "LLAMA_ARG_CACHE_TYPE_V": "q8_0"
          }
        }
      ]
    },
    {
      "id": "qwen3.6-27b-ud-q4-k-xl",
      "name": "Qwen 3.6 27B",
      "family": "qwen",
      "gguf_file": "Qwen3.6-27B-UD-Q4_K_XL.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3.6-27B-GGUF/resolve/82d411acf4a06cfb8d9b073a5211bf410bfc29bf/Qwen3.6-27B-UD-Q4_K_XL.gguf",
      "gguf_sha256": "ff6941ded525b34eb159496762c29dd0ec6e71dc31b74d57e75d871a03eec259",
      "size_bytes": 17612564704,
      "size_mb": 17613,
      "vram_required_gb": 22,
      "context_length": 65536,
      "max_context_length": 262144,
      "block_count": 64,
      "attention_head_count_kv": [0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4, 0, 0, 0, 4],
      "attention_key_length": 256,
      "attention_value_length": 256,
      "total_params_b": 26.9,
      "active_params_b": 26.9,
      "architecture": "hybrid-dense",
      "quantization": "UD-Q4_K_XL",
      "specialty": "Quality",
      "description": "Candidate for a side-by-side test against Qwen 3.5 27B on 24-32 GB GPUs; not an install default. It has the same architecture as Qwen 3.5 27B, so memory use and speed should match (about 21 GiB at 64K context). Qwen's own table scores it higher (Terminal-Bench 2.0 59.3 vs 41.6, SWE-bench Verified 77.2 vs 75.0), but those runs used thinking mode and ODS serves with reasoning off. Known risks: instruction following may be weaker, llama.cpp does not enforce tool_choice required for Qwen 3.6 with thinking off (#27767), and it repeated one tool call in a loop in 3 of 216 MMBT cells. No ODS app results yet.",
      "tokens_per_sec_estimate": 25,
      "llm_model_name": "qwen3.6-27b",
      "llama_server_image": null,
      "install_recommendation": false,
      "source_repo": "unsloth/Qwen3.6-27B-GGUF",
      "source_revision": "82d411acf4a06cfb8d9b073a5211bf410bfc29bf",
      "source_url": "https://huggingface.co/unsloth/Qwen3.6-27B-GGUF",
      "license": "apache-2.0",
      "publisher_evidence": {
        "source_url": "https://huggingface.co/Qwen/Qwen3.6-27B/blob/6a9e13bd6fc8f0983b9b99948120bc37f49c13e9/README.md",
        "baseline_model": "Qwen/Qwen3.5-27B",
        "terminal_bench_2_0": 59.3,
        "baseline_terminal_bench_2_0": 41.6,
        "swe_bench_verified": 77.2,
        "baseline_swe_bench_verified": 75.0,
        "swe_bench_pro": 53.5,
        "baseline_swe_bench_pro": 51.2,
        "independent": false,
        "note": "Same-table publisher results with thinking-mode sampling (temperature 1.0, top_p 0.95). ODS serves local models with LLAMA_ARG_REASONING=off, so read the deltas as direction, not size. The Qwen 3.6 27B card has no instruction-following row; the possible dip comes only from cross-table comparison (Qwen 3.5 27B card: IFBench 76.5), so exact-text behavior needs fleet evidence."
      }
    },
    {
      "id": "gemma4-26b-a4b-q4",
      "name": "Gemma 4 26B-A4B",
      "family": "gemma4",
      "gguf_file": "gemma-4-26B-A4B-it-UD-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/gemma-4-26B-A4B-it-GGUF/resolve/c099eb48e663fd284577b04978a94ffccb261841/gemma-4-26B-A4B-it-UD-Q4_K_M.gguf",
      "gguf_sha256": "f2c28b3dc4776931ac6f879e11f203dec637ea0f14267a86ec8f6165f63f293f",
      "size_bytes": 16947541728,
      "size_mb": 16162,
      "vram_required_gb": 19,
      "context_length": 65536,
      "max_context_length": 262144,
      "block_count": 30,
      "attention_layer_count": 5,
      "attention_head_count_kv": 2,
      "attention_key_length": 512,
      "attention_value_length": 512,
      "sliding_window": 1024,
      "sliding_window_layer_count": 25,
      "sliding_window_head_count_kv": 8,
      "sliding_window_key_length": 256,
      "sliding_window_value_length": 256,
      "recurrent_state_bytes": 0,
      "architecture_source": {
        "repo": "google/gemma-4-26B-A4B-it",
        "revision": "4d7ae4984b7db7de8f8457170b3f1a419ee76d52",
        "file": "config.json",
        "note": "Sliding-window layers hold only the window (llama.cpp b9014 caches n_swa + n_ubatch cells for them), so KV grows only on the full-attention layers."
      },
      "selection": {
        "discrete": 70,
        "unified": 70,
        "cpu": 70
      },
      "quantization": "UD-Q4_K_M",
      "specialty": "Reasoning",
      "description": "Google's large MoE model. Strong at reasoning and analysis tasks.",
      "tokens_per_sec_estimate": 50,
      "llm_model_name": "gemma-4-26b-a4b-it",
      "llama_server_image": "ghcr.io/ggml-org/llama.cpp:server-cuda-b9014@sha256:fcf285820892e7ce3218379634e3590826fc697e8b6745b9392072462e355c4f"
    },
    {
      "id": "qwen3-30b-a3b-q4",
      "name": "Qwen 3 30B-A3B",
      "family": "qwen",
      "gguf_file": "Qwen3-30B-A3B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3-30B-A3B-GGUF/resolve/d5b1d57bd0b504ac62ae6c725904e96ef228dc74/Qwen3-30B-A3B-Q4_K_M.gguf",
      "gguf_sha256": "9f1a24700a339b09c06009b729b5c809e0b64c213b8af5b711b3dbdfd0c5ba48",
      "size_mb": 18600,
      "vram_required_gb": 22,
      "context_length": 40960,
      "max_context_length": 40960,
      "app_compatibility": {
        "hermes_talk": {
          "status": "unsupported",
          "label": "Context too small for ODS Talk",
          "reason": "config.json max_position_embeddings is 40,960 with no rope_scaling, below Hermes's 65,536 minimum context.",
          "userNote": "ODS Talk needs at least 64K of context; this model supports only 40K.",
          "evidence": "https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/ad44e777bcd18fa416d9da3bd8f70d33ebb85d39/config.json"
        }
      },
      "install_recommendation": false,
      "install_recommendation_reason": "Native context is 40,960 (Qwen/Qwen3-30B-A3B config.json) and it was served at 131,072 without YaRN; the tier-map fallback now uses Qwen3.6-35B-A3B. Kept for manual activation.",
      "quantization": "Q4_K_M",
      "specialty": "Quality",
      "description": "High-quality MoE model with a 40K native context. Excellent for complex reasoning.",
      "tokens_per_sec_estimate": 55,
      "llm_model_name": "qwen3-30b-a3b",
      "llama_server_image": null
    },
    {
      "id": "gemma4-31b-q4",
      "name": "Gemma 4 31B",
      "family": "gemma4",
      "gguf_file": "gemma-4-31B-it-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/gemma-4-31B-it-GGUF/resolve/c1ac76e99d5513b141e8adde7288b85c3f9c32ec/gemma-4-31B-it-Q4_K_M.gguf",
      "gguf_sha256": "38bd64c852c4b460434cc7162fa9bdcf242faf86502581a754cb72956bb17f84",
      "size_bytes": 18323733440,
      "size_mb": 17475,
      "vram_required_gb": 24,
      "context_length": 131072,
      "selection": {
        "discrete": 80,
        "unified": 80,
        "cpu": 80
      },
      "quantization": "Q4_K_M",
      "specialty": "Quality",
      "description": "Google's flagship dense model with 131K context. Top-tier quality for enterprise GPUs.",
      "tokens_per_sec_estimate": 40,
      "llm_model_name": "gemma-4-31b-it",
      "llama_server_image": "ghcr.io/ggml-org/llama.cpp:server-cuda-b9014@sha256:fcf285820892e7ce3218379634e3590826fc697e8b6745b9392072462e355c4f"
    },
    {
      "id": "deepseek-r1-32b-q4",
      "name": "DeepSeek R1 32B",
      "family": "deepseek",
      "gguf_file": "DeepSeek-R1-Distill-Qwen-32B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF/resolve/1938d05cc893a60f37be1dc16e7465038f4fca63/DeepSeek-R1-Distill-Qwen-32B-Q4_K_M.gguf",
      "gguf_sha256": "ca171ca03554ee20cf67ad6b540610ae7eabb95af00c0abd36bb73542e140fb5",
      "size_bytes": 19851335584,
      "size_mb": 19900,
      "vram_required_gb": 36,
      "context_length": 65536,
      "max_context_length": 131072,
      "block_count": 64,
      "attention_head_count_kv": 8,
      "attention_key_length": 128,
      "attention_value_length": 128,
      "recurrent_state_bytes": 0,
      "architecture_source": {
        "repo": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
        "revision": "711ad2ea6aa40cfca18895e8aca02ab92df1a746",
        "file": "config.json"
      },
      "install_recommendation": false,
      "install_recommendation_reason": "January 2025 reasoning distill with no tool training; next in line on 24 GB. Kept for manual activation.",
      "quantization": "Q4_K_M",
      "specialty": "Reasoning",
      "description": "DeepSeek R1 distilled into Qwen 32B. The best open-source reasoning model at this size.",
      "tokens_per_sec_estimate": 25,
      "llm_model_name": "deepseek-r1-distill-qwen-32b",
      "llama_server_image": null
    },
    {
      "id": "qwen3.5-35b-a3b-q4",
      "name": "Qwen 3.5 35B-A3B",
      "family": "qwen",
      "gguf_file": "Qwen3.5-35B-A3B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3.5-35B-A3B-GGUF/resolve/bc014a17be43adabd7066b7a86075ff935c6a4e2/Qwen3.5-35B-A3B-Q4_K_M.gguf",
      "gguf_sha256": "3b46d1066bc91cc2d613e3bc22ce691dd77e6f0d33c9060690d24ce6de494375",
      "size_mb": 22000,
      "vram_required_gb": 24,
      "context_length": 131072,
      "install_recommendation": false,
      "install_recommendation_reason": "Retired for Qwen3.6-35B-A3B, which has the same layout and better same-table scores. Kept for manual activation.",
      "quantization": "Q4_K_M",
      "specialty": "Quality",
      "description": "Successor to Qwen3 30B-A3B. MoE with 131K context \u2014 top scores on coding, math, and reasoning benchmarks.",
      "tokens_per_sec_estimate": 50,
      "llm_model_name": "qwen3.5-35b-a3b",
      "llama_server_image": null
    },
    {
      "id": "qwen3.6-35b-a3b-ud-q4",
      "name": "Qwen 3.6 35B-A3B",
      "family": "qwen",
      "gguf_file": "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF/resolve/a483e9e6cbd595906af30beda3187c2663a1118c/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf",
      "gguf_sha256": "ac0e2c1189e055faa36eff361580e79c5bd6f8e76bffb4ce547f167d53e31a61",
      "size_bytes": 22134528992,
      "size_mb": 21110,
      "vram_required_gb": 24,
      "context_length": 131072,
      "max_context_length": 262144,
      "block_count": 40,
      "attention_layer_count": 10,
      "attention_head_count_kv": 2,
      "attention_key_length": 256,
      "attention_value_length": 256,
      "recurrent_state_bytes": 65863680,
      "architecture_source": {
        "repo": "Qwen/Qwen3.6-35B-A3B",
        "revision": "995ad96eacd98c81ed38be0c5b274b04031597b0",
        "file": "config.json"
      },
      "selection": {
        "discrete": 90,
        "unified": 90,
        "cpu": 0,
        "min_capacity_gib": {
          "discrete": 36
        }
      },
      "quantization": "UD-Q4_K_M",
      "specialty": "Quality",
      "description": "Current A3B MoE successor for Spark-class unified-memory hosts. Verified on DGX Spark with llama.cpp b9014.",
      "tokens_per_sec_estimate": 59,
      "llm_model_name": "qwen3.6-35b-a3b",
      "llama_server_image": null
    },
    {
      "id": "kat-coder-v2.5-dev-apex-q4",
      "name": "KAT-Coder V2.5 Dev 35B-A3B",
      "family": "kat-coder",
      "gguf_file": "KAT-Coder-V2.5-Dev-APEX-I-Balanced.gguf",
      "gguf_url": "https://huggingface.co/mudler/KAT-Coder-V2.5-Dev-APEX-GGUF/resolve/be23ff3a49eee0d5160e3fd4f5d58062160856c2/KAT-Coder-V2.5-Dev-APEX-I-Balanced.gguf",
      "gguf_sha256": "ee6e0ec15964c42ba91831d13e9709239f1b74da3dc49dd3edec4ad6aed8029f",
      "size_bytes": 25268441472,
      "size_mb": 25269,
      "vram_required_gb": 32,
      "context_length": 131072,
      "max_context_length": 262144,
      "total_params_b": 35,
      "active_params_b": 3,
      "architecture": "hybrid-moe",
      "quantization": "APEX-I-Balanced",
      "specialty": "Agentic Coding",
      "description": "Kuaishou's July 2026 SFT and RL post-train of Qwen 3.6 35B-A3B for repository-scale coding, tool discipline, and efficient agent loops. Independent 120-run evidence tied the best pass rate while using about half the stock models' input tokens.",
      "tokens_per_sec_estimate": 55,
      "llm_model_name": "kat-coder-v2.5-dev",
      "llama_server_image": null,
      "install_recommendation": false,
      "quality_evidence": {
        "source": "K. Marble 35B Coder Bakeoff",
        "source_url": "https://kmarble.dev/posts/35b-coder-bakeoff/",
        "benchmark": "Pre-registered 120-run agentic task battery",
        "score": "29/30",
        "assessed_at": "2026-07-27",
        "independent": true,
        "note": "KAT tied the best of four 35B-class models at 29/30 passes, used 12.6M input tokens versus 25.1M for stock Qwen 3.6, and had zero malformed tool-call leaks in the cited research task. The study used the MTP-grafted APEX I-Balanced sibling; this 40-block artifact removes the grafted draft head and therefore requires its own fleet validation. The study also had one harness, one GPU, five repetitions per task, and a quantization confound favoring the post-trains."
      },
      "publisher_evidence": {
        "source_url": "https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev",
        "swe_bench_verified": 69.4,
        "swe_bench_multilingual": 63.0,
        "swe_bench_pro": 45.96,
        "terminal_bench_2_1": 41.02,
        "pinch_bench": 93.43,
        "scicode": 44.2,
        "kat_code_bench": 46.21,
        "independent": false,
        "note": "Publisher-reproduced results use disclosed 256K agent harnesses and one run per evaluation. They are supporting evidence, not a substitute for fleet validation or independent reruns."
      },
      "source_evidence": {
        "model_card": "https://huggingface.co/Kwaipilot/KAT-Coder-V2.5-Dev",
        "model_revision": "7be56fe773e72b6f5ca93c1ae45d828ddb893922",
        "weights_revision": "68466af897fddf5539a9a7179b42868645457a95",
        "gguf_repository": "mudler/KAT-Coder-V2.5-Dev-APEX-GGUF",
        "gguf_revision": "be23ff3a49eee0d5160e3fd4f5d58062160856c2",
        "license": "apache-2.0",
        "license_url": "https://www.apache.org/licenses/LICENSE-2.0",
        "license_note": "The KAT checkpoint is Apache-2.0 licensed; commercial use, modification, and redistribution are permitted subject to license notice and attribution requirements.",
        "gguf_metadata_note": "The exact artifact reports qwen35moe, 256 experts with 8 active, 262K native context, and 40 model blocks with no grafted MTP draft layer. It embeds the upstream pre-fix chat template (SHA-256 e84f32a23fdda27689f868aa4a1a5621f41133e51a48d7f3efcbea2839574259), which rejects non-leading system messages; upstream fixed that template at revision 3a7d874090df0cd4399401982eca67df2c5a7e82, so exact deployed-consumer validation is required before recommendation."
      },
      "rejected_artifacts": [
        {
          "repository": "gbuzhf/KAT-Coder-V2.5-Dev-APEX-MTP-GGUF",
          "revision": "b5d04175d9c1d014d61a0c26eef987f50656d8b4",
          "file": "Kwaipilot_KAT-Coder-V2.5-Dev-APEX-MTP-I-Balanced.gguf",
          "sha256": "cce1395f842b99c147cc99025267c866aaaec5cd7b8c2c9c11d5530dc10c5e7b",
          "verdict": "runtime_incompatible",
          "reason": "The grafted 41st MTP block requires llama.cpp MTP loader support merged after the ODS production builds. ODS llama.cpp b8668 on Strix Halo/ROCm and b9014 on Tower2/CUDA both interpreted it as a normal hybrid block and failed with missing tensor 'blk.40.ssm_conv1d.weight' despite ample memory.",
          "evidence": {
            "product_sha": "333b08c4f65de6534c46a45e2fd646732629c76f",
            "harness_sha": "65a24267982811bc72e45db219e09e6e547c9628",
            "run": "2026-07-29T09-28-59Z-model-ui-product-333b08c4f65d-harness-65a242679828-hosts-tower2-strix-halo-spark-m5-mbp-strixy-kat-coder-fit-five-r1",
            "hosts": [
              "tower2",
              "strix-halo"
            ],
            "observed_at": "2026-07-29"
          }
        }
      ]
    },
    {
      "id": "deepseek-r1-70b-q4",
      "name": "DeepSeek R1 70B",
      "family": "deepseek",
      "gguf_file": "DeepSeek-R1-Distill-Llama-70B-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF/resolve/732dd974083ea5877d7b6d788b36fe7c2e5eab36/DeepSeek-R1-Distill-Llama-70B-Q4_K_M.gguf",
      "gguf_sha256": "952ff479c48ac3ece81fb6d9a5a03bfeac215e6caac780fbe91ff8cb0e05bcf3",
      "size_bytes": 42520396448,
      "size_mb": 42500,
      "vram_required_gb": 61,
      "context_length": 65536,
      "max_context_length": 131072,
      "block_count": 80,
      "attention_head_count_kv": 8,
      "attention_key_length": 128,
      "attention_value_length": 128,
      "recurrent_state_bytes": 0,
      "architecture_source": {
        "repo": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
        "revision": "b1c0b44b4369b597ad119a196caf79a9c40e141e",
        "file": "config.json"
      },
      "install_recommendation": false,
      "install_recommendation_reason": "Needs about 50.5 GiB at 32K (10 GiB of KV) and 60.5 GiB at 64K, more than the 48-62 GB hosts it was picked for; no tool training. Kept for manual activation.",
      "quantization": "Q4_K_M",
      "specialty": "Reasoning",
      "description": "The largest DeepSeek R1 distill. State-of-the-art reasoning for dual-GPU or 48GB+ setups.",
      "tokens_per_sec_estimate": 15,
      "llm_model_name": "deepseek-r1-distill-llama-70b",
      "llama_server_image": null
    },
    {
      "id": "qwen3-coder-next-q4",
      "name": "Qwen 3 Coder Next",
      "family": "qwen",
      "gguf_file": "qwen3-coder-next-Q4_K_M.gguf",
      "gguf_url": "https://huggingface.co/unsloth/Qwen3-Coder-Next-GGUF/resolve/ce09c67b53bc8739eef83fe67b2f5d293c270632/Qwen3-Coder-Next-Q4_K_M.gguf",
      "gguf_sha256": "9e6032d2f3b50a60f17ce8bf5a1d85c71af9b53b89c7978020ae7c660f29b090",
      "size_bytes": 48528320544,
      "size_mb": 48500,
      "vram_required_gb": 52,
      "context_length": 131072,
      "max_context_length": 262144,
      "block_count": 48,
      "attention_layer_count": 12,
      "attention_head_count_kv": 2,
      "attention_key_length": 256,
      "attention_value_length": 256,
      "recurrent_state_bytes": 79036416,
      "architecture_source": {
        "repo": "Qwen/Qwen3-Coder-Next",
        "revision": "a7fbcb5c0e12d62a448eaa0e260346bf5dcc0feb",
        "file": "config.json"
      },
      "selection": {
        "discrete": 100,
        "unified": 100,
        "cpu": 0
      },
      "quantization": "Q4_K_M",
      "specialty": "Code",
      "description": "Flagship coding model for enterprise GPUs and Strix Halo. 131K context for large codebases.",
      "tokens_per_sec_estimate": 30,
      "llm_model_name": "qwen3-coder-next",
      "llama_server_image": null
    },
    {
      "id": "llama4-scout-q4",
      "name": "Llama 4 Scout",
      "family": "llama4",
      "gguf_file": "Llama-4-Scout-17B-16E-Instruct-Q4_K_M-00001-of-00002.gguf",
      "gguf_url": "",
      "gguf_sha256": "",
      "gguf_parts": [
        {
          "file": "Llama-4-Scout-17B-16E-Instruct-Q4_K_M-00001-of-00002.gguf",
          "url": "https://huggingface.co/unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF/resolve/72a6853f56a66dc13a3a4b6bdc9cf7ee4c364b47/Q4_K_M/Llama-4-Scout-17B-16E-Instruct-Q4_K_M-00001-of-00002.gguf",
          "sha256": "fbe956902467171ed7c0c326e5d868771a84d46468d407abecd0f289297313f9"
        },
        {
          "file": "Llama-4-Scout-17B-16E-Instruct-Q4_K_M-00002-of-00002.gguf",
          "url": "https://huggingface.co/unsloth/Llama-4-Scout-17B-16E-Instruct-GGUF/resolve/72a6853f56a66dc13a3a4b6bdc9cf7ee4c364b47/Q4_K_M/Llama-4-Scout-17B-16E-Instruct-Q4_K_M-00002-of-00002.gguf",
          "sha256": "e7330ae14527f08e8a44a6a573b45084052b8822c4b8a5179c5cad9cd6f6f795"
        }
      ],
      "size_mb": 65300,
      "vram_required_gb": 70,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "General",
      "description": "Meta's flagship MoE with 16 experts. 131K context, strong multimodal capabilities. Needs 80GB+ VRAM.",
      "tokens_per_sec_estimate": 20,
      "llm_model_name": "llama-4-scout",
      "llama_server_image": null
    },
    {
      "id": "qwen3.5-122b-a10b-q4",
      "name": "Qwen 3.5 122B-A10B",
      "family": "qwen",
      "gguf_file": "Qwen3.5-122B-A10B-Q4_K_M-00001-of-00003.gguf",
      "gguf_url": "",
      "gguf_sha256": "",
      "gguf_parts": [
        {
          "file": "Qwen3.5-122B-A10B-Q4_K_M-00001-of-00003.gguf",
          "url": "https://huggingface.co/unsloth/Qwen3.5-122B-A10B-GGUF/resolve/51eab4d59d53f573fb9206cb3ce613f1d0aa392b/Q4_K_M/Qwen3.5-122B-A10B-Q4_K_M-00001-of-00003.gguf",
          "sha256": "467c9bd92ea518539cf75bf5a5fbfbd35e9a0b40d766ccaa67bf120e12041df3"
        },
        {
          "file": "Qwen3.5-122B-A10B-Q4_K_M-00002-of-00003.gguf",
          "url": "https://huggingface.co/unsloth/Qwen3.5-122B-A10B-GGUF/resolve/51eab4d59d53f573fb9206cb3ce613f1d0aa392b/Q4_K_M/Qwen3.5-122B-A10B-Q4_K_M-00002-of-00003.gguf",
          "sha256": "90db14846413aebdac365b57206441437cac5f7e5037d94b325f0167f902e6e7"
        },
        {
          "file": "Qwen3.5-122B-A10B-Q4_K_M-00003-of-00003.gguf",
          "url": "https://huggingface.co/unsloth/Qwen3.5-122B-A10B-GGUF/resolve/51eab4d59d53f573fb9206cb3ce613f1d0aa392b/Q4_K_M/Qwen3.5-122B-A10B-Q4_K_M-00003-of-00003.gguf",
          "sha256": "e3c24b8ebec070bb4f69ea0aca25a16531da7440cd515529953e046882901f97"
        }
      ],
      "size_mb": 76500,
      "vram_required_gb": 80,
      "context_length": 131072,
      "quantization": "Q4_K_M",
      "specialty": "Quality",
      "description": "The largest Qwen 3.5 MoE. 122B params, 10B active. Top-tier quality for multi-GPU or 96GB+ unified memory systems.",
      "tokens_per_sec_estimate": 15,
      "llm_model_name": "qwen3.5-122b-a10b",
      "llama_server_image": null
    },
    {
      "id": "qwen3.8-flash-next-nvfp4-vllm",
      "name": "Qwen 3.8 Flash Next NVFP4 (vLLM)",
      "family": "qwen",
      "runtime": "vllm",
      "tools": true,
      "vision": true,
      "hf_repo": "nvidia/Qwen3.8-Flash-Next-NVFP4",
      "revision": "fc694b54fb0174e0913e6adf86691ef85a4ead47",
      "source_url": "https://huggingface.co/nvidia/Qwen3.8-Flash-Next-NVFP4",
      "license": "other",
      "license_name": "nvidia-open-model-license",
      "size_bytes": 132734506847,
      "size_mb": 132735,
      "served_model_name": "Qwen3.8-Flash-Next-NVFP4",
      "llm_model_name": "qwen3.8-flash-next-nvfp4",
      "context_length": 131072,
      "max_context_length": 262144,
      "min_gpu_mib": 158720,
      "compute_capability": ["10.0", "10.3"],
      "reasoning": "medium",
      "thinking_template_kwargs": {
        "off": {"enable_thinking": false},
        "low": {"reasoning_effort": "low"},
        "medium": {"reasoning_effort": "medium"},
        "high": {"reasoning_effort": "xhigh"}
      },
      "tool_call_parser": "qwen3_coder",
      "reasoning_parser": "qwen3",
      "vllm_args": ["--quantization", "modelopt", "--max-model-len", "131072", "--max-num-seqs", "32", "--max-num-batched-tokens", "8192", "--enable-prefix-caching", "--enable-auto-tool-choice", "--tool-call-parser", "qwen3_coder", "--reasoning-parser", "qwen3", "--trust-remote-code", "--engram-config.cpu_offload=false", "--speculative-config.method=mtp", "--speculative-config.num_speculative_tokens=3"],
      "selection": {
        "discrete": 0,
        "unified": 0,
        "cpu": 0,
        "vllm": 90
      },
      "quantization": "NVFP4",
      "specialty": "Reasoning and coding",
      "description": "NVIDIA's NVFP4 build of Qwen 3.8 Flash Next for the vLLM engine (not llama.cpp): a mixture-of-experts model with 125B parameters, 6B active, plus n-gram embeddings and an MTP head. It reads text, images and video. It needs one NVIDIA datacenter Blackwell GPU (compute capability 10.x, such as a B200, B300 or GB300) with at least 155 GiB of memory, and 133 GB of disk. Served at 128K context with medium thinking and MTP speculative decoding. On a DGX Station GB300 it wrote about 510 tokens/s on one coding stream and read prompts at about 31,000 tokens/s.",
      "files": [
        {"path": ".gitattributes", "size": 1635, "git_blob_sha1": "a09db2ea4d1bd1fc1c09f6fca45e8e4054953535"},
        {"path": "README.md", "size": 12109, "git_blob_sha1": "f3edfc5b8f2cae8ae751162626d4c5aa3ed071a6"},
        {"path": "chat_template.jinja", "size": 8952, "git_blob_sha1": "c0c686f9c38d70d179fb7b5f5aa7530bc913dda3"},
        {"path": "config.json", "size": 30820, "git_blob_sha1": "d9df7636dc5b8bdab82655aba8ce0ebe25b22342"},
        {"path": "generation_config.json", "size": 202, "git_blob_sha1": "023756cfadf88e5bf69eefeee3e172f38c448d64"},
        {"path": "hf_quant_config.json", "size": 23001, "git_blob_sha1": "89bb1a48d2358f4b9ab7959bd4357a6d6d04020e"},
        {"path": "merges.txt", "size": 3353259, "git_blob_sha1": "a494e019ca1502219fd0128658b979e5f05ae8e8"},
        {"path": "model-00001-of-00010.safetensors", "size": 3115991696, "sha256": "63fde954be6f08b49b876f4f70a0ad0bcfee71aff7b1faa33779b6b32feca2a2"},
        {"path": "model-00002-of-00010.safetensors", "size": 10005510304, "sha256": "4dafaef62a908e49e0d92c7c2a3fa99f4d9651fdeeb0ca09f09a9b40095af4e0"},
        {"path": "model-00003-of-00010.safetensors", "size": 10005364728, "sha256": "3218ddc129258e91a721a8e81329ca8f00588332ec721bd8d2cf7d20e324bf47"},
        {"path": "model-00004-of-00010.safetensors", "size": 10006060112, "sha256": "55e2bdf6a3a1f6e65270787f65c63b2a2a56b308b54eb1fb318dfabc9b48b141"},
        {"path": "model-00005-of-00010.safetensors", "size": 10005362392, "sha256": "d3c169d3694bfba846455d01e48f047d770f78d071c5a4d22f129389411d686e"},
        {"path": "model-00006-of-00010.safetensors", "size": 10005375328, "sha256": "38f65a9d11090428e139cc60d0587e8e5433ea2ce79820883c583d7cfcc53130"},
        {"path": "model-00007-of-00010.safetensors", "size": 10005956544, "sha256": "8cf3fa05c04cb2e060963b69bb01f7a7b9f43c58435fbe380836a0457a19cd3e"},
        {"path": "model-00008-of-00010.safetensors", "size": 10005375200, "sha256": "53d1f80746aa0cc0a7c2f4836587002a8a4dd72c827432cd2e43abae3c3899e4"},
        {"path": "model-00009-of-00010.safetensors", "size": 5679113808, "sha256": "eaff6a87ece6fbfd6ad0fbaa0a5cf9ff542f8076acba38e62410b1ed55e12cc3"},
        {"path": "model-00010-of-00010.safetensors", "size": 128587536, "sha256": "0d49b0cf7c15bf9b5bd1e17403b36727315860a8f2a2e0f5a53655547084f640"},
        {"path": "model-fp8-mtp-ple.safetensors", "size": 53717551730, "sha256": "3525520c8602d850003eb1960aec0b64291dae33b83f8b65dd639a451df78823"},
        {"path": "model.safetensors.index.json", "size": 31275518, "sha256": "660414e8300728ba80062e6a3c81cb76a65e2ab136ac9e9ceb450ba0c51c3c0d"},
        {"path": "preprocessor_config.json", "size": 390, "git_blob_sha1": "2ea84a437d448ff71b08df68fdd949d5cc4ebb64"},
        {"path": "processor_config.json", "size": 1191, "git_blob_sha1": "33818c7f9e991ad735fd240209f4fa73e6c28c50"},
        {"path": "tokenizer.json", "size": 12809320, "sha256": "0997f410c57a1f4e53b09e4be8f4a172d90edd9564368fb0847030937229b9f3"},
        {"path": "tokenizer_config.json", "size": 17928, "git_blob_sha1": "5de744b3fca2129d7186979ae47c06be33903243"},
        {"path": "video_preprocessor_config.json", "size": 385, "git_blob_sha1": "3ba673a5ad7d4d13f54155ecd38b2a94a6dac8fe"},
        {"path": "vocab.json", "size": 6722759, "git_blob_sha1": "0aa0ce0658d60ac4a5d609f4eadb0e8e43514176"}
      ]
    },
    {
      "id": "glm-5.3-flash-nvfp4-vllm",
      "name": "GLM 5.3 Flash NVFP4 (vLLM)",
      "family": "glm",
      "runtime": "vllm",
      "tools": true,
      "vision": true,
      "app_compatibility": {
        "pixel_agent": {
          "status": "verified",
          "label": "Pixel verified on a DGX Station (vLLM)",
          "reason": "A fresh LLM_BACKEND=vllm install on Workstation06 (DGX Station GB300) served this model and passed Portal visitor rounds 5 and 7: a mood journal that remembers entries, built and checked in 48.6 s; a cited answer that matched GitHub; a CSV download; exact free GPU memory; a follow-up feature; a false bug report investigated without changing code; a three-page site checked page by page; and a clarifying question for an ambiguous request.",
          "evidence": "Workstation06 ~/.local/state/claude-e2e/round5-glm-vllm.json and round7-glm-vllm.json; summarized on Osmantic/ODS#8298",
          "recordedAt": "2026-10-27T22:05:00Z",
          "productSha": "43143c59c1bb031394432af11c120d927e113d5b",
          "hostScope": ["workstation06"]
        }
      },
      "hf_repo": "nvidia/GLM-5.3-Flash-NVFP4",
      "revision": "da920bb0b9f4a06727223a349e55468e38352348",
      "source_url": "https://huggingface.co/nvidia/GLM-5.3-Flash-NVFP4",
      "license": "mit",
      "size_bytes": 204476278199,
      "size_mb": 204476,
      "served_model_name": "GLM-5.3-Flash-NVFP4",
      "llm_model_name": "glm-5.3-flash-nvfp4",
      "context_length": 131072,
      "max_context_length": 1048576,
      "min_gpu_mib": 244736,
      "compute_capability": ["10.0", "10.3"],
      "reasoning": "medium",
      "thinking_template_kwargs": {
        "off": {"reasoning_effort": "low"},
        "low": {"reasoning_effort": "low"},
        "medium": {"reasoning_effort": "high"},
        "high": {}
      },
      "tool_call_parser": "glm47",
      "reasoning_parser": "glm47",
      "vllm_args": ["--max-model-len", "131072", "--kv-cache-dtype", "fp8", "--max-num-seqs", "32", "--max-num-batched-tokens", "8192", "--enable-prefix-caching", "--enable-auto-tool-choice", "--tool-call-parser", "glm47", "--reasoning-parser", "glm47", "--trust-remote-code", "--speculative-config.method=mtp", "--speculative-config.num_speculative_tokens=5", "--no-enable-flashinfer-autotune"],
      "selection": {
        "discrete": 0,
        "unified": 0,
        "cpu": 0,
        "vllm": 100
      },
      "quantization": "NVFP4",
      "specialty": "Quality",
      "description": "NVIDIA's NVFP4 build of GLM-5.3 Flash for the vLLM engine (not llama.cpp): a mixture-of-experts model with 320B parameters, 18B active. It reads text, images and video. It needs one NVIDIA datacenter Blackwell GPU (compute capability 10.x) with at least 239 GiB of memory, such as a GB300, and 204 GB of disk. Served at 128K context with an FP8 KV cache, medium thinking and MTP speculative decoding. On a DGX Station GB300 it wrote about 390 tokens/s on one coding stream and read prompts at about 24,000 tokens/s.",
      "files": [
        {"path": ".gitattributes", "size": 1635, "git_blob_sha1": "a09db2ea4d1bd1fc1c09f6fca45e8e4054953535"},
        {"path": ".quant_summary.txt", "size": 616642, "git_blob_sha1": "388d3ad297bef5d95516a0400716d27af6e0eec7"},
        {"path": "README.md", "size": 10235, "git_blob_sha1": "08fa91a54c57d22b41740736017c83a1477d5d00"},
        {"path": "chat_template.jinja", "size": 10644, "git_blob_sha1": "fb94d40d72a2eefa0c7140cb66d941656be16969"},
        {"path": "config.json", "size": 16770, "git_blob_sha1": "c76d04db0f678a8f7e8b1b6a27a97b6050b7b340"},
        {"path": "generation_config.json", "size": 215, "git_blob_sha1": "46d04e68002526d394949f63d65f93ac4ac8abeb"},
        {"path": "hf_quant_config.json", "size": 8076, "git_blob_sha1": "cbc53ce3eec597abe641fbeaac4aa50fd01e4fd6"},
        {"path": "model-00001-of-00033.safetensors", "size": 8945680672, "sha256": "5c7ee3e2a0bd7452b7ccea70bd18f7221676a709a3d4c050d1457fa7d7359125"},
        {"path": "model-00002-of-00033.safetensors", "size": 8945659984, "sha256": "7afab1480693d8a37faba2f092b0626b121f6fce9241478ad95cf857145ddc41"},
        {"path": "model-00003-of-00033.safetensors", "size": 8945740832, "sha256": "f184e8b75db437579fa571d4693d6fba7a2806e531d3b8e05d13b1dbc0a90142"},
        {"path": "model-00004-of-00033.safetensors", "size": 5985863344, "sha256": "dd08e62f3f86dad5ac784e68eb956fa20a18bab72d81337d3058af281ce18c93"},
        {"path": "model-00005-of-00033.safetensors", "size": 5985909512, "sha256": "bc763b5eec6911a7750b6e37d7f499d6226535792be5c7e7a169d1abb09875d6"},
        {"path": "model-00006-of-00033.safetensors", "size": 5915365198, "sha256": "c7d9597fad87af3b54dbcf994be2b65cd99fc458e392e34f22eeac4af502e4fc"},
        {"path": "model-00007-of-00033.safetensors", "size": 5915359020, "sha256": "3d0f68feb7fe4cee3bd87a464c845843594e0039124941c2a89e7c1ce9da9e2b"},
        {"path": "model-00008-of-00033.safetensors", "size": 5915364798, "sha256": "5c4d7657a3d6fe137b9de14a917d84e511fc9b206fe76bd943d1e53deedba857"},
        {"path": "model-00009-of-00033.safetensors", "size": 5915358812, "sha256": "4770820b433f6ca50c7d30f0b2eaa9df6fd7c095769f51ee46aa519c37e5f7d3"},
        {"path": "model-00010-of-00033.safetensors", "size": 5915358494, "sha256": "be92de14713dc46c0b0c8c87f1c02395b2eeb6059087c69d0a39d01dee34edd8"},
        {"path": "model-00011-of-00033.safetensors", "size": 5915358332, "sha256": "a3294966995e1a81525c6a9efdeb6a27c599a4d04b4e3a5342331298a527d0be"},
        {"path": "model-00012-of-00033.safetensors", "size": 5915083008, "sha256": "4d6f4b1a555377bfa773981a4c0d84a86ee782501613a18a3e86caefbace7f66"},
        {"path": "model-00013-of-00033.safetensors", "size": 5915639642, "sha256": "d1d5c86fa3566eb099f99677550f44191c77faa935491d6d2c5a7c337797eaf9"},
        {"path": "model-00014-of-00033.safetensors", "size": 5915084432, "sha256": "fd6c81655a572a2462dea772bed6fd5114017f12185a619ce693f6b476be5524"},
        {"path": "model-00015-of-00033.safetensors", "size": 5915642946, "sha256": "6ff8daa650fcfacd509e15be48f4c986ff7dae081a874fcb4c1ee67d3a66487a"},
        {"path": "model-00016-of-00033.safetensors", "size": 5915084544, "sha256": "286c1a635749f6364c5b4bbab0c7a9eb0916f9421b17f6ade618a5478b25be58"},
        {"path": "model-00017-of-00033.safetensors", "size": 5915653746, "sha256": "7be78c2466c0c5d4472c1405ff8631b1da1f0dc059e878d3f6f75bf670514b27"},
        {"path": "model-00018-of-00033.safetensors", "size": 5915084536, "sha256": "b5755440acc16e3a935ab60106a99bc9db6cc03d033786e7502e2010a27ff6aa"},
        {"path": "model-00019-of-00033.safetensors", "size": 5915653908, "sha256": "2385071e3600654014011755c97a46e1abe1d2aa988e6fdb7aa2bd4c751b9419"},
        {"path": "model-00020-of-00033.safetensors", "size": 5915368968, "sha256": "8597c39572caa35b1a5e4896ed7c83eb12350ca862dce875071ec2e2d70fec98"},
        {"path": "model-00021-of-00033.safetensors", "size": 5915366420, "sha256": "cf2923ceb37fd1dcb2772f16aa6d459e472dea3d36275f5c5580781646f27f60"},
        {"path": "model-00022-of-00033.safetensors", "size": 5915362564, "sha256": "8f27fac338e9960367a58f0a6ba3681148d52bf38d563d183cac5ab7092b16a6"},
        {"path": "model-00023-of-00033.safetensors", "size": 5915370484, "sha256": "04bbd26640bdf145da1c412aa4c01ef2d0103b2d7e8a2846f207175be461be5d"},
        {"path": "model-00024-of-00033.safetensors", "size": 5915084584, "sha256": "2dc1b2fdf6ba9816545ce9036c9561cd293f71c887d3d3d1590e86900548fb75"},
        {"path": "model-00025-of-00033.safetensors", "size": 5915652952, "sha256": "12a327e0f52b18e7feae2936ebacf33f7a5eff11c96038a06d2547f5653f0838"},
        {"path": "model-00026-of-00033.safetensors", "size": 5915084552, "sha256": "0eb6db2777bebcf9ea34e9b639a8497ffb0d7c520e00a4e0a0b6331cfb18ac2d"},
        {"path": "model-00027-of-00033.safetensors", "size": 5915652344, "sha256": "e220abe4abb6a4798b1d16fdb7d4f6121b6c6e3e8f8d9f71243ba957a7f9542d"},
        {"path": "model-00028-of-00033.safetensors", "size": 5915364684, "sha256": "04b44953c3a5019c4da79260b9abf39bc6fa2b9414f4ede20fea400325aaab3d"},
        {"path": "model-00029-of-00033.safetensors", "size": 5915373524, "sha256": "a705d67e2580270744f26135fc46fd062ea59df6260b45cd628bc2b8a74f6ab2"},
        {"path": "model-00030-of-00033.safetensors", "size": 5915364844, "sha256": "60c0e12a793693702303e7d0ebf16e7619fbbc7a45fe2e9e6d038459053f8e1f"},
        {"path": "model-00031-of-00033.safetensors", "size": 5915373956, "sha256": "5d3a07314659b264b56d84385e89abb5a0199ce944a7cb33eab556832d100075"},
        {"path": "model-00032-of-00033.safetensors", "size": 5915364628, "sha256": "5126cddeb35041953ad0375d926081f7d78b0b279460fa8d1afcdfbe36bca3d1"},
        {"path": "model-00033-of-00033.safetensors", "size": 5915373132, "sha256": "c6fa7921452627ad9d7913875d78496c60082f71b05f943b5dddb3318f0771c2"},
        {"path": "model.safetensors.index.json", "size": 16291474, "sha256": "26765b2601fd246ef361cfb9f5e10f9fb291a59e05ad0a109062f3a4747c7fd1"},
        {"path": "processor_config.json", "size": 909, "git_blob_sha1": "3ec2a058b2cdec556eca0f621e8c5164eec213c2"},
        {"path": "tokenizer.json", "size": 20217442, "sha256": "19e773648cb4e65de8660ea6365e10acca112d42a854923df93db4a6f333a82d"},
        {"path": "tokenizer_config.json", "size": 761, "git_blob_sha1": "e375fa0a4e16660ce0e2026e39374d35dad4c079"}
      ]
    },
    {
      "id": "qwen3.6-35b-a3b-nvfp4-vllm",
      "name": "Qwen 3.6 35B A3B NVFP4 (vLLM)",
      "family": "qwen",
      "runtime": "vllm",
      "tools": true,
      "vision": true,
      "hf_repo": "nvidia/Qwen3.6-35B-A3B-NVFP4",
      "revision": "1355db6a052410cfd62085d94b58866fd0f2c3c5",
      "source_url": "https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4",
      "license": "apache-2.0",
      "size_bytes": 23462477857,
      "size_mb": 23462,
      "served_model_name": "Qwen3.6-35B-A3B-NVFP4",
      "llm_model_name": "qwen3.6-35b-a3b-nvfp4",
      "context_length": 131072,
      "max_context_length": 262144,
      "min_gpu_mib": 51200,
      "gpu_memory_utilization": 0.5,
      "compute_capability": ["12.0"],
      "reasoning": "medium",
      "thinking_template_kwargs": {
        "off": {"enable_thinking": false},
        "low": {"enable_thinking": true},
        "medium": {"enable_thinking": true},
        "high": {"enable_thinking": true}
      },
      "tool_call_parser": "qwen3_coder",
      "reasoning_parser": "qwen3",
      "vllm_args": ["--max-model-len", "131072", "--max-num-seqs", "32", "--max-num-batched-tokens", "8192", "--enable-prefix-caching", "--enable-auto-tool-choice", "--tool-call-parser", "qwen3_coder", "--reasoning-parser", "qwen3", "--trust-remote-code", "--speculative-config.method=mtp", "--speculative-config.num_speculative_tokens=3"],
      "selection": {
        "discrete": 0,
        "unified": 0,
        "cpu": 0,
        "vllm": 80
      },
      "quantization": "NVFP4",
      "specialty": "Speed",
      "description": "NVIDIA's NVFP4 build of Qwen 3.6 35B A3B for the vLLM engine (not llama.cpp): a mixture-of-experts model with 35B parameters, 3B active, plus an MTP head. It reads text, images and video. It needs one RTX PRO 6000-class GPU (Blackwell, compute capability 12.0) with at least 50 GiB of memory, and 24 GB of disk. vLLM takes half of that GPU, so image generation and voice still fit beside it on a workstation's only GPU. Served at 128K context with thinking on and MTP speculative decoding. On a DGX Station's RTX PRO 6000 it wrote about 540 tokens/s on one coding stream (llama.cpp's Qwen 3.6 Q4: 250) and read prompts at about 36,000 tokens/s (10,300).",
      "files": [
        {"path": ".gitattributes", "size": 1635, "git_blob_sha1": "aa7aacd0134a92c3c1943fdecc75cd8b7420cce6"},
        {"path": ".quant_summary.txt", "size": 4752286, "git_blob_sha1": "b97d531c8e2bd7cbd1c42cbcb49f029090010251"},
        {"path": "README.md", "size": 9936, "git_blob_sha1": "0ecb54c1554791cedb4615b7ced45826b9817d19"},
        {"path": "chat_template.jinja", "size": 7764, "git_blob_sha1": "a8755d827c0a7b614c246c4060dfd58ab352a8ff"},
        {"path": "config.json", "size": 58110, "git_blob_sha1": "ec1f82c006fd5db6d0b5fa1a59e16fbda9a5c1dd"},
        {"path": "configuration.json", "size": 58, "git_blob_sha1": "d24dba949ee1fe70cc810e4c4709a0bddf4e06ba"},
        {"path": "generation_config.json", "size": 202, "git_blob_sha1": "023756cfadf88e5bf69eefeee3e172f38c448d64"},
        {"path": "hf_quant_config.json", "size": 35085, "git_blob_sha1": "8deefe44cc69cfab994141ea6b9493b03c864061"},
        {"path": "model-00001-of-00003.safetensors", "size": 10006877608, "sha256": "07141c2db92e47bc08777132cd1a0323faf300eab3a7d7c111bc2bf075fda050"},
        {"path": "model-00002-of-00003.safetensors", "size": 10003595752, "sha256": "6dea9c759a0f941cf9e1cc1501216b0e107966a66c90425050e212efbd053f02"},
        {"path": "model-00003-of-00003.safetensors", "size": 3413864960, "sha256": "9758875fc55e49561165f4a44342b654c123c3e25c6811b34abb83553fb1a164"},
        {"path": "model.safetensors.index.json", "size": 13726227, "sha256": "d67403a4e9793c0ba8a136baf14b3b76ec7b32c822267978084895e07ebd8a3e"},
        {"path": "preprocessor_config.json", "size": 390, "git_blob_sha1": "2ea84a437d448ff71b08df68fdd949d5cc4ebb64"},
        {"path": "tokenizer.json", "size": 12807982, "sha256": "5f9e4d4901a92b997e463c1f46055088b6cca5ca61a6522d1b9f64c4bb81cb42"},
        {"path": "tokenizer_config.json", "size": 16718, "git_blob_sha1": "28d96ff303d1d20350185caf4bf037045916ed35"},
        {"path": "video_preprocessor_config.json", "size": 385, "git_blob_sha1": "3ba673a5ad7d4d13f54155ecd38b2a94a6dac8fe"},
        {"path": "vocab.json", "size": 6722759, "git_blob_sha1": "0aa0ce0658d60ac4a5d609f4eadb0e8e43514176"}
      ]
    }
  ]
}