Nanbeige4.1-3B: A Compact and Versatile Model

Nanbeige LLM Lab has released Nanbeige4.1-3B, an open-source language model with 3 billion parameters. The main goal of this project is to demonstrate that a small model can achieve high performance in several key areas, including complex reasoning, alignment with human preferences, and agentic capabilities.

Key Features

  • Advanced Reasoning: Nanbeige4.1-3B is capable of solving complex problems through sustained and coherent reasoning, achieving significant results in challenging benchmarks such as LiveCodeBench-Pro, IMO-Answer-Bench, and AIME 2026 I.
  • Alignment with Human Preferences: The model demonstrates strong alignment with human preferences, achieving a score of 73.2 on Arena-Hard-v2 and 52.21 on Multi-Challenge, surpassing larger models.
  • Agentic Capabilities: In addition to chat tasks, Nanbeige4.1-3B natively supports deep-search functionalities and achieves remarkable results in tasks such as xBench-DeepSearch and GAIA.
  • Extended Context and Prolonged Reasoning: The model supports contexts up to 256,000 tokens, enabling deep-search with hundreds of tool calls and single-pass reasoning for complex problems requiring over 100,000 tokens.

The model is available on Hugging Face. A technical report is forthcoming.

For those evaluating on-premise deployments, there are trade-offs discussed in detail on /llm-onpremise.