logo

Jan-nano: A Compact AI Model for Deep Research

June 17, 2025•By LLM Hard Drive Store
Jan-nano: A Compact AI Model for Deep Research
AIDeep ResearchJan-nanoMenlo Research

Jan-nano: A Compact AI Model for Deep Research

Key Points

  • Jan-nano is a compact 4-billion parameter AI model for deep research, likely outperforming larger models with MCP integration.
  • It runs offline via the Jan platform, ensuring privacy and user control, with over 3 million downloads.
  • Research suggests it excels in real-time web searches and deep research, supported by SimpleQA benchmark scores.
  • The evidence leans toward Jan-nano being efficient, needing moderate hardware like 8GB RAM for basic use.

Introduction to Jan-nano

Jan-nano, developed by Menlo Research, is a small yet powerful AI model designed for deep research tasks. It integrates with the Model Context Protocol (MCP) to connect with external tools, enhancing its research capabilities. This article explores its features, performance, and how to use it, based on recent announcements and community discussions.

About Menlo Research and Jan

Menlo Research is an open R&D lab focused on AI tools that boost human-machine collaboration. Their product, Jan, is an open-source, offline AI assistant available on GitHub, with over 3 million downloads. It aims to provide privacy and control, contrasting with cloud-based AI services.

Features and Performance

Jan-nano operates offline, ensuring data privacy, and is optimized for MCP, enabling real-time web searches and deep research. It scored 80.7 on the SimpleQA benchmark with MCP, reportedly outperforming larger models like DeepSeek-v3-671B (78.2), according to community discussions on Reddit.

Getting Started

To use Jan-nano, install the Jan Beta from Jan's website, download the model from the Hub, enable MCP with a Serper API key, and start researching via the chat interface. Check the official documentation for detailed setup.

Background and Context

On June 17, 2025, Menlo Research announced Jan-nano via an X post, highlighting its ability to outperform DeepSeek-v3-671B using MCP. This 4-billion parameter model is part of the Jan ecosystem, an open-source AI assistant designed to run offline, ensuring user privacy. The GitHub repository for Jan, located at GitHub, reveals it has over 3 million downloads, underscoring its popularity as a privacy-focused alternative to cloud-based AI like ChatGPT.

Menlo Research, described on their website (Menlo Research), is an R&D lab aiming to advance human-machine collaboration through open-source tools. Their vision, detailed at Jan's About Page, is to create a generally intelligent agent that operates autonomously, evolving computers from user-operated to self-driving, as inspired by Bill Gates' notes on AI agents (Gates Notes).

What is Jan-nano?

Jan-nano is a compact language model fine-tuned from Qwen3-4B, as noted on its Hugging Face page (Menlo/Jan-nano). It is designed for deep research tasks, optimized for MCP servers, which facilitate integration with research tools and data sources. The model size is 4.02B parameters, using BF16 tensor type, making it efficient for local deployment. Its documentation, available at Setup, Usage & FAQ, emphasizes its role in tool-augmented research.

The X post video, lasting 55.63 seconds, showcased Jan-nano's interface, demonstrating tasks like summarizing AI industry news and financial reports, with buttons for "Real-time web search" and "Deep research." A frame at 33.38 seconds explicitly stated, "Jan Nano by Menlo Beats DeepSeek-V3-671B using MCP. Check description to run the beta version," reinforcing its performance claims.

Features and Capabilities

Jan-nano's key features include:

  • Compact Size: Requiring only 8GB RAM for iQ4_XS quantization and 12GB VRAM for Q8, as per the documentation, it runs efficiently on standard hardware.
  • MCP Integration: MCP, an open standard by Anthropic (Anthropic MCP Introduction), standardizes AI connections to external tools, enabling Jan-nano to perform real-time web searches and deep research. It acts like a "USB-C for AI apps," as described on Wikipedia, facilitating seamless data access.
  • Local Operation: Supported by Jan Beta, available at Jan Beta Docs, it ensures privacy by running offline, with no personal information requested, as noted in the GitHub repository's trust and safety notes.
  • Research Capabilities: It handles tasks like summarizing news, financial modeling, and competitor comparisons, as seen in the video frames, with recommended sampling parameters (Temperature: 0.7, Top-p: 0.8, Top-k: 20, Min-p: 0) for optimal performance.

System requirements include macOS 13.6+ with 8GB RAM for 3B models, Windows 10+ with GPU support, and Linux compatibility, detailed in the GitHub repository. Training costs are estimated under $100 on RunPod using H200, with hardware like 8xA6000 for training and 4xA6000 for inferencing.

Performance and Benchmarks

Jan-nano's performance was evaluated on the SimpleQA benchmark using MCP, as per its Hugging Face model card. Community discussions on Reddit reveal it scored 80.7, outperforming DeepSeek-671B (78.2), ChatGPT-4.5 (62.5), and Claude-3.7-Sonnet (50.0), among others. The evaluation used an agentic setup, allowing the model to choose tools freely, reflecting real-world performance. The post notes, "Our original goal was to build a super small model that excels at using search tools to extract high-quality information," aiming for performance improvements to 85-90%.

The Hugging Face page for Jan-nano-gguf (Menlo/Jan-nano-gguf) mentions excellent function calling and tool integration, ideal for local environments. Recommended GGUF quantizations are Q8 for best performance and iQ4_XS for limited VRAM, avoiding Q4_0 and Q4_K_M due to degradation.

Conclusion and Future Outlook

Jan-nano exemplifies Menlo Research's vision of user-owned, efficient AI, aligning with Jan's goal of autonomous, privacy-focused agents. Its performance on SimpleQA suggests it could serve as a self-hosted Perplexity alternative, with plans for further improvements. As of June 17, 2025, a full technical report is pending, but current data indicates Jan-nano is a promising tool for researchers seeking local, powerful AI solutions.

Key Citations