Skip to content

Open-source platform for 3D AI agents. Turn text or a photo into a rigged, animated GLB avatar, give it an LLM brain, memory and a wallet, and embed it anywhere with one web component. Remote MCP server, x402 payments, Three.js, browser-native.

docs(llm): document the Nemotron Ultra, Lightning, Kilo and keyless LLM7 rungs with live timings

nirholasd0bdcf4eb8c0
diff --git a/ARCHITECTURE.md b/ARCHITECTURE.md
index a3c6b3a87..9b39c5e4b 100644
--- a/ARCHITECTURE.md
+++ b/ARCHITECTURE.md
@@ -2966,7 +2966,7 @@ The codebase references **~260 distinct `process.env.*` keys** across `api/`; `a
| **Core infra** | `DATABASE_URL` · `UPSTASH_REDIS_REST_URL/_TOKEN` · `UPSTASH_CACHE_REST_URL/_TOKEN` · `UPSTASH_MANAGEMENT_API_KEY` · `UPSTASH_EMAIL` · `KV_REST_API_URL/_TOKEN` · `S3_ENDPOINT/_BUCKET/_PUBLIC_DOMAIN/_ACCESS_KEY_ID/_SECRET_ACCESS_KEY` · `JWT_SECRET` · `JWT_KID` · `WALLET_ENCRYPTION_KEY` · `INTERNAL_API_KEY` · `CRON_SECRET` · `CHROMIUM_PACK_URL` · `APP_ORIGIN` · `PUBLIC_APP_ORIGIN` · `DEFAULT_APP_ORIGIN` · `STOREFRONT_ORIGIN` · `SERVER_DOMAIN` · `NODE_ENV` · `VERCEL_ENV` |
| **RPC & chains** | `SOLANA_RPC_URL` · `SOLANA_RPC_URL_DEVNET` · `MAINNET_RPC_URL` · `BASE_RPC_URL` · `BASE_SEPOLIA_RPC_URL` · `SEPOLIA_RPC_URL` · `EVM_RPC_URL` · `RPC_URL_8453` · `RPC_URL_84532` · `CLUB_BASE_RPC_URL` · `A2A_REPUTATION_RPC_URL` · `HELIUS_API_KEY` · `ETHERSCAN_API_KEY` |
| **$THREE & treasuries** | `THREE_TOKEN_MINT` · `THREE_TOKEN_DECIMALS` · `THREE_TREASURY_WALLET` · `THREE_REWARDS_WALLET` · `THREE_BURN_ADDRESS` · `THREE_QUOTE_SECRET` · `THREE_QUOTE_TTL_S` · `THREE_PRIZE_PAYOUT_KEY` · `THREE_ALLOWANCE_DELEGATE_SECRET_KEY_B64` · `THREE_WS_MARK_ENFORCE` · `PLATFORM_TREASURY_KEYPAIR` · `TREASURY_KEYPAIR` · `COIN_TREASURY_SECRET_KEY_B64` · `COIN_CREATOR_SECRET_KEY_B64_*` · `CLUB_SOLANA_TREASURY_SECRET_KEY_B64` · `CLUB_EVM_TREASURY_PRIVATE_KEY` |
-| **LLM / AI** | `ANTHROPIC_API_KEY` · `OPENAI_API_KEY` · `GROQ_API_KEY` · `OPENROUTER_API_KEY` · `OPENROUTER_FALLBACK_KEYS` · `NVIDIA_API_KEY` · `NVIDIA_A2F_FUNCTION_ID` · `NVIDIA_ASR_FUNCTION_ID` · `NVIDIA_COSMOS_INVOKE_URL` · `WATSONX_API_KEY/_URL/_PROJECT_ID/_SPACE_ID/_MODEL_ID/_GUARDIAN_MODEL_ID/_API_VERSION` · `WATSONX_ORCHESTRATE_URL/_API_KEY/_AGENT` · `DASHSCOPE_API_KEY` · `MODELSCOPE_API_KEY` · `LIVEPEER_API_KEY` · `VOYAGE_API_KEY` · `AIXBT_API_KEY/_API_BASE/_ENABLED` · `ELEVENLABS_API_KEY` |
+| **LLM / AI** | `ANTHROPIC_API_KEY` · `OPENAI_API_KEY` · `GROQ_API_KEY` · `OPENROUTER_API_KEY` · `OPENROUTER_FALLBACK_KEYS` · `NVIDIA_API_KEY` · `NVIDIA_A2F_FUNCTION_ID` · `NVIDIA_ASR_FUNCTION_ID` · `NVIDIA_COSMOS_INVOKE_URL` · `KILO_API_KEY` (optional, keyless lane) · `LLM7_API_KEY` (optional, keyless lane) · `WATSONX_API_KEY/_URL/_PROJECT_ID/_SPACE_ID/_MODEL_ID/_GUARDIAN_MODEL_ID/_API_VERSION` · `WATSONX_ORCHESTRATE_URL/_API_KEY/_AGENT` · `DASHSCOPE_API_KEY` · `MODELSCOPE_API_KEY` · `LIVEPEER_API_KEY` · `VOYAGE_API_KEY` · `AIXBT_API_KEY/_API_BASE/_ENABLED` · `ELEVENLABS_API_KEY` |
| **x402 / payments** | `X402_FACILITATOR_URL(_BASE/_SOLANA)` · `X402_FACILITATOR_TOKEN(_BASE/_SOLANA)` · `X402_CDP_FACILITATOR_URL` · `X402_PAY_TO(_BASE/_BSC/_SOLANA)` · `X402_ASSET_ADDRESS_BASE/_BSC/_ARBITRUM` · `X402_ASSET_MINT(_SOLANA)` · `X402_FEE_PAYER_SOLANA` · `X402_ACCEPT_THREE_SOLANA` · `X402_THREE_AMOUNT_SOLANA` · `X402_EVM_NETWORKS` · `X402_MAX_PER_CALL/_HOUR/_DAY_ATOMIC` · `X402_MAX_AMOUNT_REQUIRED` · `X402_IDEMPOTENCY_TTL_SECONDS` · `X402_BUILDER_CODE_APP/_WALLET` · `X402_PRICE_<SLUG>` (per-endpoint overrides) · `CDP_API_KEY_ID/_SECRET` · `BSC_OPERATOR_KEY` · `PLATFORM_FEE_BPS` |
| **x402 self-facilitator / ring** | `X402_SELF_FACILITATOR_ENABLED` · `X402_FEE_PAYER_SECRET_BASE58` · `X402_TREASURY_SECRET_BASE58` · `X402_SEED_SOLANA_SECRET_BASE58` · `X402_SELF_FACILITATOR_PAYTO_ALLOWLIST` · `X402_SPONSOR_SOL_FLOOR_LAMPORTS` · `X402_SELF_FAC_MAX_CU_LIMIT/_CU_PRICE/_PRIORITY_LAMPORTS` · `X402_RING_SELF_PAY` · `X402_PRICE_RING_SETTLE` · `X402_RING_TREASURY_BUFFER_ATOMIC` · `X402_RING_MIN_SWEEP_ATOMIC` · `X402_EXTERNAL_ENABLED` · `X402_AUTONOMOUS_DAILY_CAP_ATOMIC` · `X402_VOLUME_PER_RUN_CAP_ATOMIC` · `X402_VOLUME_BATCH_PER_RUN` |
| **Circulation engine** | `CIRCULATION_ENABLED` · `CIRCULATION_TREASURY_SECRET` · `CIRCULATION_NETWORK` · `CIRCULATION_POOL_TARGET` · `CIRCULATION_GROWTH_PER_TICK` · `CIRCULATION_ACTIONS_PER_TICK` · `CIRCULATION_EVM_TREASURY_SECRET` · `CIRCULATION_EVM_CHAIN_ID` (removed: `CIRCULATION_REAL_SELLER_DEMAND`, `_DAILY_CAP`) |
@@ -3391,7 +3391,7 @@ Tiers: **CRITICAL** = `req()` in `api/_lib/env.js` (cold-start throw) or prod-fa
**10. Granite Guardian gate (`api/_lib/granite-guardian.js`).** 12-risk taxonomy, each an independent watsonx classifier pass; default input risks `[jailbreak,harm,violence,unethical_behavior,social_bias]` run concurrently. Each pass recovers a real unsafe-probability by softmaxing Yes/No logprobs (`logprobs:true, top_logprobs:5`); `decide` → allow/review/block at `FLAG_THRESHOLD=0.55`. `governSend` adds a model-independent `sendCapUsd` (default $25). **Hash-chained audit:** stores only SHA-256 of content + decision + `prev` hash; `verifyAuditChain` re-derives from `GENESIS_HASH`.
-**11. LLM failover (`api/_lib/chat-models.js`, `api/chat.js`).** `DEFAULT_PROVIDER_ORDER = [groq, openrouter, nvidia, sambanova, mistral, zai, anthropic, openai, grok]`: free lanes first, paid as last-resort backstop (prod paid keys routinely 401/over-quota). The shared utility chain (`api/_lib/llm.js providerChain()`) is wider still: groq → cerebras → openrouter `:free` → nvidia → sambanova → mistral → zai → cloudflare (Workers AI) → ovh (keyless) → gemini → vertex-gemini (anchor) → pollinations (keyless) → llm7 (keyless) → siliconflow → groq#instant → paid tail; every keyed free rung is skipped when its env var is unset, and the three keyless rungs (OVH, Pollinations, LLM7) are unconditional. `MODEL_CATALOG` annotates `{provider, tools, moderationGated}`; `usableModels` drops tool-incapable or moderation-gated models from auto-chains. `api/chat.js` skips providers in health cooldown (recent 429/5xx/timeout) + unconfigured keys, bounds the loop by `TOTAL_BUDGET_MS=25000` + `MAX_FALLBACK_ATTEMPTS=4` (per-call abort at `min(15000, remaining)`); auth/billing failures cool the whole provider; exhausted-within-budget → clean **503 + Retry-After:20**.
+**11. LLM failover (`api/_lib/chat-models.js`, `api/chat.js`).** `DEFAULT_PROVIDER_ORDER = [groq, openrouter, nvidia, sambanova, mistral, zai, anthropic, openai, grok]`: free lanes first, paid as last-resort backstop (prod paid keys routinely 401/over-quota). The shared utility chain (`api/_lib/llm.js providerChain()`) is wider still: groq → groq#120b → cerebras → openrouter `:free` → nvidia (Nemotron Super) → nvidia:ultra (Nemotron Ultra) → sambanova → mistral → zai → cloudflare (Workers AI) → huggingface → kilo (keyless, Nemotron Ultra then Super) → ovh (keyless) → gemini → vertex-gemini (anchor) → pollinations (keyless) → llm7 (keyless) → siliconflow → nvidia:lightning → groq#instant → paid tail; every keyed free rung is skipped when its env var is unset, and the keyless rungs (Kilo, OVH, Pollinations, LLM7) are unconditional. `/api/cron/free-model-audit` checks the OpenRouter, NVIDIA, and Kilo pins against their live catalogs every 6 hours. `MODEL_CATALOG` annotates `{provider, tools, moderationGated}`; `usableModels` drops tool-incapable or moderation-gated models from auto-chains. `api/chat.js` skips providers in health cooldown (recent 429/5xx/timeout) + unconfigured keys, bounds the loop by `TOTAL_BUDGET_MS=25000` + `MAX_FALLBACK_ATTEMPTS=4` (per-call abort at `min(15000, remaining)`); auth/billing failures cool the whole provider; exhausted-within-budget → clean **503 + Retry-After:20**.
### Specifications, Protocols & Platform Constants
@@ -3421,7 +3421,7 @@ Hard, load-bearing specifics from `specs/` and `docs/`. (The $THREE hold-to-acce
**x402 wire constants** — `X402_VERSION=2`, scheme `exact`. USDC: Solana `EPjFWdd5AufqSSqeM2qN1xzybapC8G4wEGGkZwyTDt1v`, Base `0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913`, Arbitrum `0xaf88d065e77c8cC2239327C5EDb3A432268e5831`. Facilitators: PayAI `https://facilitator.payai.network`, Coinbase CDP `https://api.cdp.coinbase.com/platform/v2/x402` (Base+Arbitrum). CAIP-2: Base `eip155:8453`, Arbitrum `eip155:42161`, BSC `eip155:56`, Solana `solana:5eykt4UsFv8P8NJdTREpY1vzqKqZKvdp`. $THREE accepted as second Solana settlement asset (`acceptThree:true`, opt-in `X402_ACCEPT_THREE_SOLANA`).
-**AI model IDs.** NVIDIA (`NVIDIA_API_KEY`, NIM base `integrate.api.nvidia.com/v1`): text→3D `microsoft/trellis` (steps 10-50, prompt ≤77 chars), text→image `black-forest-labs/flux.1-schnell`, default LLM `meta/llama-3.3-70b-instruct`, vision `nvidia/nemotron-nano-12b-v2-vl`→`meta/llama-3.2-11b-vision-instruct`, embeddings `nvidia/nv-embedqa-e5-v5` (1024-dim, ≤512 tok)→`baai/bge-m3`, rerank `nvidia/rerank-qa-mistral-4b`, safety `nvidia/llama-3.1-nemoguard-8b-content-safety`→`meta/llama-guard-4-12b`, TTS `magpie-tts-multilingual`. IBM watsonx: chat `ibm/granite-3-8b-instruct`, embeddings `ibm/granite-embedding-278m-multilingual`, forecast `ibm/granite-ttm-{512|1024|1536}-96-r2`, vision `ibm/granite-vision-3-2-2b`, governance `ibm/granite-guardian-3-8b`; Identity Firewall similarity ≥93% = block, 86-93% = review. i18n provider ladder: vertex (default, billed to GCP credits) → gemini (CJK) → groq → openrouter → nvidia → mistral → openai → anthropic → threews (we-pay proxy); every run also falls over to OpenRouter mid-chunk; glossary `doNotTranslate` includes `$THREE`, contract address, `x402`, `USDC`, `Solana`, `IBM watsonx.ai`.
+**AI model IDs.** NVIDIA (`NVIDIA_API_KEY`, NIM base `integrate.api.nvidia.com/v1`): text→3D `microsoft/trellis` (steps 10-50, prompt ≤77 chars), text→image `black-forest-labs/flux.1-schnell`, chat LLMs `nvidia/nemotron-3-super-120b-a12b` → `nvidia/nemotron-3-ultra-550b-a55b` → `nvidia/nemotron-3.5-lightning-30b-a3b` (thinking off; `meta/llama-3.x` reached end of life on NIM 2026-08-26), vision `nvidia/nemotron-nano-12b-v2-vl`→`meta/llama-3.2-11b-vision-instruct`, embeddings `nvidia/nv-embedqa-e5-v5` (1024-dim, ≤512 tok)→`baai/bge-m3`, rerank `nvidia/rerank-qa-mistral-4b`, safety `nvidia/llama-3.1-nemoguard-8b-content-safety`→`meta/llama-guard-4-12b`, TTS `magpie-tts-multilingual`. IBM watsonx: chat `ibm/granite-3-8b-instruct`, embeddings `ibm/granite-embedding-278m-multilingual`, forecast `ibm/granite-ttm-{512|1024|1536}-96-r2`, vision `ibm/granite-vision-3-2-2b`, governance `ibm/granite-guardian-3-8b`; Identity Firewall similarity ≥93% = block, 86-93% = review. i18n provider ladder: vertex (default, billed to GCP credits) → gemini (CJK) → groq → openrouter → nvidia → mistral → openai → anthropic → threews (we-pay proxy); every run also falls over to OpenRouter mid-chunk; glossary `doNotTranslate` includes `$THREE`, contract address, `x402`, `USDC`, `Solana`, `IBM watsonx.ai`.
**Pump platform fee** (`docs/pump-platform-fee.md`) — `PUMP_PLATFORM_FEE_BPS` default 0 (off), hard-capped 500 (5%); recipient `PUMP_PLATFORM_FEE_WALLET` → treasury fallback; buy fee added on top, sell fee from proceeds.
diff --git a/data/changelog.json b/data/changelog.json
index b5730d25a..c5169b2d0 100644
--- a/data/changelog.json
+++ b/data/changelog.json
@@ -29044,6 +29044,13 @@
"summary": "Agent pages, agent profiles and coin pages now carry an Earned card: the creator fees an agent's coins have earned on pump.fun, in SOL with dollars beside it, how much has been claimed and how much is still waiting, a per-coin breakdown, and every claim linked to its transaction on Solscan. Skill sales and hires by other agents count too. The figures refresh every 30 minutes, each card says how old its numbers are and how they are calculated, and developers can read the same data from the public earnings API or over MCP. Automatic fee claiming for autonomous launches also works again.",
"tags": ["feature", "fix"],
"link": "/launches"
+ },
+ {
+ "date": "2026-10-08",
+ "title": "Chat and agents stay up through more AI outages",
+ "summary": "Every AI answer on three.ws, from chat to agent brains, the tutor and the fact-checker, now has five more free models to fall back on when the usual ones are busy. NVIDIA's largest open model, Nemotron 3 Ultra, now answers right behind Nemotron 3 Super, with the compact Nemotron 3.5 Lightning further down the line. A second, independent route to the same Nemotron models works with no account at all, and a Gemma 4 model is back in the mix as well. The free NVIDIA and gateway models are now checked against their providers every six hours too, so the team hears the moment one is switched off instead of finding out from a failed answer.",
+ "tags": ["improvement"],
+ "link": "/chat"
}
]
}
diff --git a/docs/nvidia-models.md b/docs/nvidia-models.md
index 4815263b5..a32974375 100644
--- a/docs/nvidia-models.md
+++ b/docs/nvidia-models.md
@@ -32,7 +32,7 @@ Because it is one key for everything, a deployment either has the whole NVIDIA l
| --- | --- | --- | --- |
| **Text → 3D** | `microsoft/trellis` | `api/_providers/nvidia.js`, `api/_lib/forge-tiers.js` | ✅ |
| **Text → image** | `black-forest-labs/flux.1-dev` | `api/_mcp3d/text-to-image.js` | ✅ |
-| **LLM (default lane)** | `nvidia/nemotron-3-super-120b-a12b` (compact rung: `nvidia/nemotron-3.5-lightning-30b-a3b`) | `api/_lib/llm.js`, `api/_lib/chat-models.js` | ✅ |
+| **LLM (default lane)** | `nvidia/nemotron-3-super-120b-a12b`, then `nvidia/nemotron-3-ultra-550b-a55b` (step-down rung: `nvidia/nemotron-3.5-lightning-30b-a3b`) | `api/_lib/llm.js`, `api/_lib/chat-models.js` | ✅ |
| **LLM (model garden)** | Nemotron 3 Super 120B / Ultra 550B / 3.5 Lightning 30B, DeepSeek V4 Pro, Kimi K2.6, MiniMax M2.7 | `api/brain/chat.js` | ✅ |
| **Vision / VLM** | `meta/llama-3.2-11b-vision-instruct`, then free OpenRouter `:free` VLM rungs | `api/_lib/vision.js` | ✅ |
| **Embeddings** | `nvidia/nemotron-3-embed-1b` | `api/_lib/embeddings.js` (and `api/agents/_id/embed.js`, which delegates to it) | ✅ |
@@ -86,11 +86,13 @@ NVIDIA NIM hosts 100+ open-weight chat models behind the one key, all OpenAI-com
### 3a. The default production lane
-**Model:** `nvidia/nemotron-3-super-120b-a12b` (with `nvidia/nemotron-3.5-lightning-30b-a3b` as the compact Nemotron rung)
+**Models:** `nvidia/nemotron-3-super-120b-a12b`, then `nvidia/nemotron-3-ultra-550b-a55b`, with `nvidia/nemotron-3.5-lightning-30b-a3b` as the compact step-down rung
**Source:** [api/_lib/llm.js](../api/_lib/llm.js), [api/_lib/chat-models.js](../api/_lib/chat-models.js).
The platform's general LLM helper runs a **free-first ladder: Groq → Cerebras → OpenRouter → NVIDIA NIM**, followed by further keyless free rungs (OVH, Gemini, Pollinations) and only at the very end a paid backstop (Anthropic/OpenAI). Every free rung was re-pointed on 2026-08-27, and every NVIDIA-routed id again on 2026-09-10 after a sweep of `GET /v1/models` on the live account found most of them answering `410 Gone`. A dead id is worse than a missing one: a fallback chain treats a 410 like a rate limit and moves on, so nothing logged an outage while each dead entry burned one of the chain's few retry slots. The rungs the providers still serve: Groq now runs `qwen/qwen3.8-27b` (instant tier `openai/gpt-oss-20b`), Cerebras keeps `llama-3.3-70b`, and the NVIDIA rung runs Nemotron 3 Super 120B, an NVIDIA MoE on an independent provider (the opt-in compact rung moved off `nemotron-3-nano-30b-a3b`, which NIM stopped serving, to `nemotron-3.5-lightning-30b-a3b`), so an outage on the other free lanes still answers here. Both NVIDIA rungs are called with thinking disabled (`chat_template_kwargs: { enable_thinking: false }`) so a chat turn answers instead of reasoning out loud. Both are tool/function-calling capable, so they are eligible for tool-required requests.
+Since 2026-10-08 the shared chain runs all three Nemotron models on the one key, as separate rungs: Super 120B, then Ultra 550B right behind it, then Lightning 30B in the step-down group near the tail. Each NIM model is its own function with its own queue and its own retirement date, and NVIDIA sizes the free rate limit by model, so one model throttling, queueing, or answering 410 no longer takes the NVIDIA lane down with it. The third-party models the same key lists (DeepSeek V4.1 Flash, Kimi K3, GLM 5.3, Gemma 4) were probed the same day and each queued past 60 seconds on the free tier, so none of them is a chain rung. The same Nemotron Ultra and Super models are also reachable keylessly through the Kilo Code gateway, which the chain uses as a separate rung on Kilo's own capacity. `/api/cron/free-model-audit` checks every pinned NIM id against the live catalog every 6 hours and pages ops when one is retired. Full chain: [free-llm-providers.md](free-llm-providers.md).
+
This lane powers the platform's built-in AI surfaces — chat, embedded site widgets, the tutor, the fact-checker, persona tools, agent-to-agent talk, the transaction explainer — all of which lead with the free providers and only fall through to a paid model if every free lane fails.
### 3b. The Brain model garden
diff --git a/docs/ops/llm-lanes.md b/docs/ops/llm-lanes.md
index e8b0550ca..14a00e5a2 100644
--- a/docs/ops/llm-lanes.md
+++ b/docs/ops/llm-lanes.md
@@ -24,16 +24,20 @@ production are routinely dead and a chain that depends on them fails:
| 2 | Cerebras `llama-3.3-70b` | `CEREBRAS_API_KEY` | free | not configured in prod |
| 3 | OpenRouter `:free` routes, one rung per key | `OPENROUTER_API_KEY`, `OPENROUTER_FALLBACK_KEYS` | free | **model id was dead 2026-09-04**: `openai/gpt-oss-20b:free` was retired and 404'd all five rungs at once (fixed, now `google/gemma-4-31b-it:free`). The account is separately capped at 1,000 free-model requests/DAY across every key (they share one owner), and was at 0 remaining when measured. |
| 4 | NVIDIA NIM `nvidia/nemotron-3-super-120b-a12b` | `NVIDIA_API_KEY` | free | **serving** (200, 1.8s-10.5s on a full-size prompt, 2026-09-04). Carried the whole platform on 2026-09-04 when Groq was token-capped, OpenRouter was dead, and both paid anchors were on billing holds. |
+| 4b | NVIDIA NIM `nvidia/nemotron-3-ultra-550b-a55b` | `NVIDIA_API_KEY` | free | **serving** (200, 1.8s through `llmComplete`, 2026-10-08). Added 2026-10-08. A separate NIM function on the same key, so it survives the Super function being throttled, queued, or retired. |
| 5 | SambaNova `Meta-Llama-3.3-70B-Instruct` | `SAMBANOVA_API_KEY` | free | added 2026-08-05; skipped when the key is unset |
| 6 | Mistral `mistral-small-latest` (Experiment tier) | `MISTRAL_API_KEY` | free | added 2026-08-05; skipped when the key is unset |
| 7 | Z.AI `glm-4.7-flash` | `ZAI_API_KEY` | free | added 2026-08-05; skipped when the key is unset |
| 8 | Cloudflare Workers AI `@cf/meta/llama-3.3-70b-instruct-fp8-fast` | `CLOUDFLARE_ACCOUNT_ID` + `CLOUDFLARE_AI_API_TOKEN` | free | added 2026-08-05; skipped when either is unset |
+| 8b | Hugging Face Inference Providers `meta-llama/Llama-3.3-70B-Instruct` | `HF_TOKEN` (+ `HF_FALLBACK_TOKENS`) | free (monthly included credit) | configured in prod; 402s in ~100ms once a token's credit is spent |
+| 8c | Kilo Code `nvidia/nemotron-3-ultra-550b-a55b:free`, then `nvidia/nemotron-3-super-120b-a12b:free` | none (keyless) | free | **serving** (200, Ultra 2.0s and Super 0.6s through `llmComplete`, 2026-10-08). Added 2026-10-08. About 200 req/hr anonymously, on Kilo's own NVIDIA capacity. Sends `reasoning: {enabled: false}`. |
| 9 | OVH AI Endpoints `Meta-Llama-3_3-70B-Instruct` | none (keyless) | free | **serving** (200, ~1.2s, 2026-08-02) |
| 10 | Gemini Flash-Lite (AI Studio) | `GEMINI_API_KEY` | free | not configured in prod |
| 11 | **Vertex Gemini Flash** | GCP service account | GCP credits | **dead**: 403 `Lightning dunning decision is deny for project` since 2026-08-27, re-measured 2026-09-04. A billing hold on the whole GCP project, not IAM. This is the chain's designed anchor, so while it is down the free rungs above it are load-bearing rather than best-effort. |
| 12 | Pollinations `openai-fast` | none (keyless) | free | **serving** (200, ~3.7s, 2026-08-02) |
-| 13 | LLM7.io `gemini-3.1-flash-lite` | `LLM7_API_KEY` | free | **dead without a key**: llm7.io retired the anonymous tier it was added on, and every unauthenticated call answers 401 `invalid_api_key` (measured 2026-09-02, the `unused` token its docs used to accept included). Key-gated since, so a deployment without the key skips it. Free key at https://dash.llm7.io/#/api-keys |
+| 13 | LLM7.io `gemma4:31b` | none (keyless); `LLM7_API_KEY` optional | free | **serving keyless** (200, 1.1s through `llmComplete`, 2026-10-08). The anonymous tier answered 401 on 2026-09-02 and reopened for LLM7's `turbo` models; the old pin `gemini-3.1-flash-lite` moved to the paid `pro` tier. Keyless again since 2026-10-08. |
| 14 | SiliconFlow `Qwen/Qwen3-8B` | `SILICONFLOW_API_KEY` | free | added 2026-08-05; skipped when the key is unset |
+| 14b | NVIDIA NIM `nvidia/nemotron-3.5-lightning-30b-a3b` | `NVIDIA_API_KEY` | free | **serving** (200, 0.7s through `llmComplete`, 2026-10-08). Added 2026-10-08 as a step-down: a third NIM function on the primary key. |
| 15 | Groq `openai/gpt-oss-20b` | `GROQ_API_KEY` | free | serving (a third separate per-model quota) |
| 16 | Vertex Claude | GCP service account + `VERTEX_CLAUDE_ENABLED=1` | GCP credits | **off and unentitled** (see below) |
| 17 | Anthropic first-party | `ANTHROPIC_API_KEY` | paid | **absent** (no key anywhere) |
@@ -43,11 +47,14 @@ production are routinely dead and a chain that depends on them fails:
The 2026-08-05 widening (SambaNova, Mistral, Z.AI, Cloudflare, LLM7,
SiliconFlow) added six independent free quota pools; each is documented with
-its tier limits in `docs/free-llm-providers.md`. LLM7 joined that round as a
-keyless rung and is no longer one (row 13), so the keyless floor a zero-env
-deployment falls to is OVH and Pollinations, two rungs rather than three. Both
-answered 429 from this workspace's shared egress IP on 2026-09-02, so treat the
-keyless floor as best-effort: it is a real lane, not a substitute for a key.
+its tier limits in `docs/free-llm-providers.md`. The 2026-10-08 round added
+the Nemotron Ultra and Lightning rungs on the existing NIM key, the keyless
+Kilo Code gateway, and made LLM7 keyless again, so the keyless floor a
+zero-env deployment falls to is now five rungs across four providers: Kilo
+(Ultra, then Super), OVH, Pollinations, and LLM7. OVH and Pollinations both
+answered 429 from this workspace's shared egress IP on 2026-09-02, so treat any
+single keyless rung as best-effort: the floor is real because it is wide, not
+because any one rung is a substitute for a key.
Rungs 1 to 15 are not a degradation path any more. They are production. Every
one of them is covered by a transport-level failover test