Small model, big claim

A 3B model that beats 9B rivals

Nanbeige4.2-3B says it can outscore models three times its size on agent benchmarks. Here is how.

The problem

Small models usually fail agent tasks

Local developers have had to choose between a slow small model or paying per call to a bigger hosted one.

The numbers

63.6 on SWE-Bench Verified

Nanbeige4.2-3B beats Qwen3.5-9B at 53.1 and Gemma4-12B at 44.2, using a Looped Transformer trained on 28 trillion tokens.

Why it works

Layers get reused, not stacked

Hidden states pass through the same layer stack twice, adding depth and capacity without adding new parameters.

The catch

Scores are self-reported so far

No independent lab has reproduced the SWE-Bench or GDPval results yet. Community testing will decide if they hold up.

Try it now

Runs today on Ollama and llama.cpp

Read the full story on the architecture, benchmarks and open-source Apache 2.0 release.