Menlo launches Jan-Nano, a 4B agentic model purpose-built for deep research with MCP
GregoryDiamos · x · 2026-09-09
Menlo Research released Jan-Nano, a compact 4-billion-parameter non-thinking LLM designed and trained for deep research tasks.
- Natively optimized to work with MCP servers as a tool-augmented research agent
- Evaluated on SimpleQA with an MCP-based benchmark methodology to reflect real-world tool-augmented performance
- Runs locally via the open-source Jan app or vLLM (with --enable-auto-tool-choice and the hermes tool-call parser); recommended sampling: temperature 0.7, top-p 0.8, top-k 20
- Weights and a technical report are public; commenters link the approach to Karpathy's "cognitive core" idea—models using tool calls instead of memorizing knowledge
More from coding & agent
- Semantic Cache Verification Measured: 30-40ms on the Hot Path vs 1.7s for an LLM Judge — Reasonable_Royal_621 · 2026-09-09
- HyperFrames hits #1 on GitHub as Claude Code-made launch video clocks 3M views in 2 days — toolstelegraph · 2026-09-09
- AI assistants may need their own email addresses as users balk at full inbox access — _AustinCalvert_ · 2026-09-09
- Dimillian's dev tool adds worktree creation on any branch after user requests — Dimillian · 2026-09-09
- Dev built a custom procedural drawing engine via Astra — all Evergrow art is drawn in code — Dimillian · 2026-09-09
- Agents researching agents: Prime Intellect's open stack eases RL environments — seanwbren · 2026-09-09