Introducing Jan-Nano: A Compact Agentic Language Model for Deep Research

Menlo Research has unveiled Jan-Nano, a streamlined 4-billion parameter (4.02B) language model purpose-built for deep research tasks and tool-augmented workflows. Built on the Qwen3 architecture, Jan-Nano excels at integrating with external data sources and research tools via the Model Context Protocol (MCP) (huggingface.co, en.wikipedia.org).

Overview

Jan-Nano is designed to run efficiently on MCP servers, enabling two-way, secure connections between the model and a variety of research datasets or tools. This “agentic” approach lets Jan-Nano fetch live context—such as document contents or API data—and incorporate it into its reasoning, effectively turning the model into a research assistant that can call upon external resources as needed (huggingface.co).

What Is the Model Context Protocol (MCP)?

The Model Context Protocol (MCP) is an open standard introduced by Anthropic in November 2024 that standardizes how AI models connect to external systems—think of it as a universal “USB-C” connector for language models. MCP defines JSON-RPC 2.0 interfaces and encourages secure, bidirectional communication, allowing models like Jan-Nano to execute functions, retrieve real-time data, and operate beyond their static training corpus (en.wikipedia.org).

Evaluation

In our MCP-based benchmarking, Jan-Nano was put through the SimpleQA evaluation suite. Despite its compact size, it demonstrated strong factual accuracy and answer consistency compared to larger, non-tool-augmented models, validating its effectiveness as a research-focused assistant (huggingface.co).

Running Jan-Nano Locally

You can deploy Jan-Nano on your machine using VLLM and the open-source Jan (beta build) interface:

vllm serve Menlo/Jan-nano \
  --host 0.0.0.0 --port 1234 \
  --enable-auto-tool-choice \
  --tool-call-parser hermes \
  --chat-template ./qwen3_nonthinking.jinja

Recommended sampling parameters to balance creativity and accuracy:

• Temperature: 0.7
• Top-p: 0.8
• Top-k: 20
• Min-p: 0

For detailed setup instructions and troubleshooting, visit the official documentation: https://menloresearch.github.io/ (huggingface.co).

Use Cases

• Academic Research: Automate literature reviews by fetching paper abstracts and summarizing findings.
• Data Analysis: Interface with local databases or cloud storage to retrieve datasets and generate insights.
• Tool-Driven Workflows: Chain API calls or code execution steps for reproducible research pipelines.

Jan-Nano represents a new wave of agentic models that blend compact footprints with powerful, context-aware capabilities, empowering researchers to work faster and smarter.