A post-training upgrade of Nanbeige4-3B. The report title is "A Small General Model that Reasons, Aligns, and Acts". There is no new pretraining: the model is initialized from Nanbeige4-3B-Base. SFT extends the context from 64K to 256K, and the post-training adds alignment and agentic skills to the reasoning. The report says it can sustain deep-search sessions of up to 600 tool-call turns. It became the lab's breakout release, with 1,193 HuggingFace likes.

Self-reported results:

  • Reasoning: GPQA 83.8, AIME 2026 I 87.4, LiveCodeBench-V6 76.9 and Humanity's Last Exam 12.6.
  • General and tool use: Arena-Hard-v2 73.2, BFCL-V4 56.5 and τ²-Bench 48.6.
  • Deep search: xBench-DeepSearch-2505 75, GAIA (text) 69.9 and BrowseComp 19.1.

Artificial Analysis lists the model but shows no Intelligence Index v5 score for it.

Model Details

Architecture DENSE
Parameters 4B
Context window 262,144
License Apache 2.0

Benchmark Scores

Benchmark Score Mode
GPQA 83.8 —
LiveCodeBench v6 76.9 —
Arena-Hard-v2 73.2 —
xBench-DeepSearch-2505 75 —
BrowseComp 19.1 —

Paper

small-modelsreasoningagentsopen-weight

Related