Skip to content

Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS

feat: refresh local LLM catalog with Qwen3.6 27B, Magistral Small 24B, and Nemotron 3.5 Lightning

[._.]/ Adam Eivybb01ce14b2ee
diff --git a/server/lib/localLlmCatalog.js b/server/lib/localLlmCatalog.js
index fb31e3b66d..9126207ba6 100644
--- a/server/lib/localLlmCatalog.js
+++ b/server/lib/localLlmCatalog.js
@@ -422,6 +422,21 @@ export const LOCAL_LLM_CATALOG = [
ollama: 'gpt-oss:20b',
lmstudio: 'lmstudio-community/gpt-oss-20b-GGUF'
},
+ {
+ key: 'magistral-small-24b',
+ name: 'Magistral Small 24B',
+ category: 'reasoning',
+ recommendedFor: ['reasoning', 'multilingual'],
+ params: '24B',
+ size: '14 GB',
+ family: 'magistral',
+ description: "Mistral AI's open-weights reasoning model with explicit thinking mode, tools, and strong multilingual capabilities for complex multi-step reasoning.",
+ repository: 'mistralai/Magistral-Small-2509',
+ capabilities: ['chat', 'reasoning', 'tools', 'multilingual'],
+ context: 40960,
+ ollama: 'magistral:24b',
+ lmstudio: 'lmstudio-community/Magistral-Small-2509-GGUF'
+ },
{
key: 'gemma3-27b-it',
name: 'Gemma 3 27B IT',
@@ -511,6 +526,21 @@ export const LOCAL_LLM_CATALOG = [
ollamaAliases: ['hf.co/unsloth/Qwen3.8-27B-GGUF:Q4_K_M'],
lmstudio: 'unsloth/Qwen3.8-27B-GGUF'
},
+ {
+ key: 'qwen3.6-27b',
+ name: 'Qwen3.6 27B',
+ category: 'general',
+ recommendedFor: ['general', 'coding', 'reasoning', 'vision', 'multilingual'],
+ params: '27B',
+ size: '18 GB',
+ family: 'qwen',
+ description: "Alibaba's dense 27B multimodal model with native vision, tools, thinking preservation, and a 262K context window — strong general, coding, and reasoning capability.",
+ repository: 'Qwen/Qwen3.6-27B',
+ capabilities: ['chat', 'code', 'reasoning', 'tools', 'vision', 'multilingual'],
+ context: 262144,
+ ollama: 'qwen3.6:27b',
+ lmstudio: 'lmstudio-community/Qwen3.6-27B-GGUF'
+ },
{
key: 'gemma4-26b-a4b',
name: 'Gemma 4 26B-A4B',
@@ -664,18 +694,19 @@ export const LOCAL_LLM_CATALOG = [
lmstudio: 'lmstudio-community/gemma-4-31B-it-GGUF'
},
{
- key: 'nemotron-3-nano-30b-a3b',
- name: 'Nemotron 3 Nano 30B-A3B',
+ key: 'nemotron-3.5-lightning',
+ name: 'Nemotron 3.5 Lightning 30B-A3B',
category: 'reasoning',
- recommendedFor: ['reasoning'],
+ recommendedFor: ['reasoning', 'coding'],
params: '30B / 3B active',
- size: '24 GB',
+ size: '25 GB',
family: 'nemotron',
- description: "NVIDIA's agentic MoE with a 1M-token context — by far the longest window here, for whole-corpus review rather than a single manuscript. Only 3B active, so it stays fast.",
- capabilities: ['chat', 'reasoning', 'tools'],
+ description: "NVIDIA's hybrid Mamba-Transformer agentic MoE with a 1M-token context — ultra-fast long-context reasoning, coding, and tool calling with only 3B active parameters.",
+ repository: 'nvidia/Nemotron-3.5-Lightning-30B-A3B',
+ capabilities: ['chat', 'reasoning', 'code', 'tools'],
context: 1048576,
- ollama: 'nemotron-3-nano:30b',
- lmstudio: 'lmstudio-community/NVIDIA-Nemotron-3-Nano-30B-A3B-GGUF'
+ ollama: 'nemotron-3.5-lightning:30b',
+ lmstudio: 'lmstudio-community/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF'
},
{
key: 'llama3.3-70b',
@@ -1117,7 +1148,11 @@ const RETIRED_MODEL_MAPPINGS = [
{ ollama: 'phi3', lmstudio: 'lmstudio-community/Phi-3.1-mini-128k-instruct-GGUF' },
{ ollama: 'deepseek-r1', lmstudio: 'lmstudio-community/DeepSeek-R1-Distill-Qwen-7B-GGUF' },
{ ollama: 'minicpm-v', lmstudio: 'openbmb/MiniCPM-V-2_6-gguf' },
- { ollama: 'smollm2', lmstudio: 'lmstudio-community/SmolLM2-1.7B-Instruct-GGUF' }
+ { ollama: 'smollm2', lmstudio: 'lmstudio-community/SmolLM2-1.7B-Instruct-GGUF' },
+ {
+ ollama: 'nemotron-3-nano:30b',
+ lmstudio: 'lmstudio-community/NVIDIA-Nemotron-3-Nano-30B-A3B-GGUF'
+ }
];
// Strip only the implicit `:latest` tag (`llama3.2:latest` → `llama3.2`) and
diff --git a/server/lib/localLlmCatalog.test.js b/server/lib/localLlmCatalog.test.js
index 44bf32ce27..13dac842e3 100644
--- a/server/lib/localLlmCatalog.test.js
+++ b/server/lib/localLlmCatalog.test.js
@@ -214,6 +214,9 @@ describe('localLlmCatalog', () => {
expect(ollama.find((m) => m.key === 'gemma4-e4b')?.id).toBe('gemma4:e4b');
expect(ollama.find((m) => m.key === 'translategemma-12b')?.id).toBe('translategemma:12b');
expect(ollama.find((m) => m.key === 'aya-expanse-8b')?.id).toBe('aya-expanse:8b');
+ expect(ollama.find((m) => m.key === 'qwen3.6-27b')?.id).toBe('qwen3.6:27b');
+ expect(ollama.find((m) => m.key === 'magistral-small-24b')?.id).toBe('magistral:24b');
+ expect(ollama.find((m) => m.key === 'nemotron-3.5-lightning')?.id).toBe('nemotron-3.5-lightning:30b');
});
});
@@ -240,18 +243,27 @@ describe('localLlmCatalog', () => {
.toEqual({ targetId: 'lmstudio-community/gemma-4-E4B-it-GGUF', exact: true });
expect(mapModelToBackend('ollama', 'translategemma:12b', 'lmstudio'))
.toEqual({ targetId: 'lmstudio-community/TranslateGemma-12B-IT-GGUF', exact: true });
+ expect(mapModelToBackend('ollama', 'qwen3.6:27b', 'lmstudio'))
+ .toEqual({ targetId: 'lmstudio-community/Qwen3.6-27B-GGUF', exact: true });
+ expect(mapModelToBackend('ollama', 'magistral:24b', 'lmstudio'))
+ .toEqual({ targetId: 'lmstudio-community/Magistral-Small-2509-GGUF', exact: true });
+ expect(mapModelToBackend('ollama', 'nemotron-3.5-lightning:30b', 'lmstudio'))
+ .toEqual({ targetId: 'lmstudio-community/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-GGUF', exact: true });
});
it('still maps models retired from the suggested-install catalog', () => {
// Retiring an entry from LOCAL_LLM_CATALOG must not downgrade an existing
// install's migrate to a guessed stem — RETIRED_MODEL_MAPPINGS keeps it exact.
expect(getCatalog('ollama').some((m) => m.key === 'llama3.2')).toBe(false);
+ expect(getCatalog('ollama').some((m) => m.key === 'nemotron-3-nano-30b-a3b')).toBe(false);
expect(mapModelToBackend('ollama', 'llama3.2:latest', 'lmstudio'))
.toEqual({ targetId: 'lmstudio-community/Llama-3.2-3B-Instruct-GGUF', exact: true });
expect(mapModelToBackend('lmstudio', 'lmstudio-community/Llama-3.2-3B-Instruct-GGUF', 'ollama'))
.toEqual({ targetId: 'llama3.2', exact: true });
expect(mapModelToBackend('ollama', 'qwen2.5vl:32b', 'lmstudio'))
.toEqual({ targetId: 'lmstudio-community/Qwen2.5-VL-32B-Instruct-GGUF', exact: true });
+ expect(mapModelToBackend('ollama', 'nemotron-3-nano:30b', 'lmstudio'))
+ .toEqual({ targetId: 'lmstudio-community/NVIDIA-Nemotron-3-Nano-30B-A3B-GGUF', exact: true });
});
it('best-effort derives an Ollama name for an unknown LM Studio model', () => {
diff --git a/server/lib/localModelHeuristics.js b/server/lib/localModelHeuristics.js
index 911d790ca7..ba5c696686 100644
--- a/server/lib/localModelHeuristics.js
+++ b/server/lib/localModelHeuristics.js
@@ -232,7 +232,7 @@ const EDITORIAL_FAMILY_RANK = [
'llama', // Llama 3.x instruct
'gemma', // Gemma 2/3/4 instruct
'glm', // GLM-4.x — strong prose, tight instruction-following
- 'mixtral', 'ministral', 'mistral', // Mistral family
+ 'mixtral', 'ministral', 'magistral', 'mistral', // Mistral family
'granite', // IBM Granite — constrained, clean structured output
'olmo', // AI2 Olmo — fully-open instruct
'command-r-plus', 'command-r', 'command', // Cohere Command — capable but chatty/RAG-tuned
diff --git a/server/lib/systemCapabilities.js b/server/lib/systemCapabilities.js
index 1ca9abb7a5..cf2dec1ee6 100644
--- a/server/lib/systemCapabilities.js
+++ b/server/lib/systemCapabilities.js
@@ -59,6 +59,8 @@ const LOCAL_LLM_HARDWARE_REQUIREMENTS = Object.freeze({
'olmo-3.1-32b': Object.freeze({ minMemoryGb: 32 }),
'gemma4-31b': Object.freeze({ minMemoryGb: 64 }),
'nemotron-3-nano-30b-a3b': Object.freeze({ minMemoryGb: 32 }),
+ 'nemotron-3.5-lightning': Object.freeze({ minMemoryGb: 32 }),
+ 'qwen3.6-27b': Object.freeze({ minMemoryGb: 32 }),
'qwen3.5-122b-a10b': Object.freeze({ minMemoryGb: 96 }),
'qwen3-coder-30b': Object.freeze({ minMemoryGb: 32 }),
'devstral-small-2-24b': Object.freeze({ minMemoryGb: 32 }),