Jan-nano

Jan-nano is a specialized 4B-parameter model from Menlo Research, introduced in June 2025 (arXiv 2506.22760). Its design philosophy is explicit: instead of storing encyclopedic knowledge in parameters, it is engineered to be excellent at finding information. Fine-tuned from Qwen3-4B with a multi-stage RLVR framework using DAPO preference optimization against a local RAG server, it learns concise tool invocation and reaches a reported 83.2% on the SimpleQA benchmark with MCP integration.

With a 128K context window and a footprint that fits consumer hardware, Jan-nano is built for the Jan ecosystem's local-first research workflows: document analysis, multi-document synthesis and agentic search. Weights are openly available on Hugging Face in standard and 128K variants. Check the model card for license terms before commercial use.