Nanbeige4.2-3B beats 9B, 12B rivals on SWE-Bench and GDPval

A model card that just landed on Hugging Face claims a 3 billion non-embedding parameter model is beating open rivals two to four times its size on agent and code benchmarks. The Nanbeige4.2-3B page reports 63.6 on SWE-Bench Verified against Qwen3.5-9B's 53.1 and Gemma4-12B's 44.2, plus 74.3 on GDPval rubrics against 61.9 and 68.5 respectively. On reasoning the story is the same shape: 87.4 on GPQA-Diamond and 82.8 on HMMT-Feb-2026, both ahead of the two larger baselines.

The lever the authors point to is a Looped Transformer that reuses layers to get more capacity out of the same parameter count, paired with what the card calls LoopSplit, mHC with depth attention, and concatenated n-gram embeddings. Training combines supervised fine-tuning on synthetic environments with reinforcement learning that mixes outcome and process rewards. The whole thing ships under Apache 2.0, with day-one recipes for vLLM, SGLang, llama.cpp, and Ollama, which is what turns this into a local-assistant story rather than a hosted-API one.

For anyone shipping an on-device copilot, the math shifts if a 3B model can carry tool-use, code editing, and reasoning workloads that used to demand something in the 9B to 12B class. The card frames it explicitly around local personal assistant deployments, and gives quantization guidance down to Q4_K_M for llama.cpp so it fits on modest hardware.

The honest caveat is that every number here comes from the authors' own model card and has not been independently reproduced. Compact models also tend to look strong on per-task benchmarks while breaking down on long-horizon workflows the tables cannot capture. What the reporting doesn't give you is training data provenance, compute budget, or a comparison against the frontier hosted agents that still set the ceiling on real-world reliability.

If the scores hold up when other evaluators re-run them, the more interesting question is whether looped-transformer tricks generalize to other open families or turn out to be a Nanbeige-specific recipe. Cheaper agents running where the data lives is the direction worth watching.