Menlo Research has introduced Jan-nano — a compact model based on Qwen3 with only 4 billion parameters, specifically designed for deep research.
This model product is ideally suited for working with the Model Context Protocol (MCP), being fully fine-tuned and optimized for seamless integration.
On the SimpleQA platform (which utilizes agent technologies and MCP), Jan-nano achieves a score of 80.7 — an impressive result for such a small model.
The model operates through Jan — open-source software alternative to ChatGPT that can be run directly on your computer without cloud services.
Jan-nano is an excellent example of how small-sized models can compete with larger ones, thanks to a smart architecture and targeted training.