Solves code and math problems and calls tools while keeping pace with models twice its size

You paste in a broken function, a multi-step math question, or a request that needs a tool call, and it thinks through the problem before it replies. It is built for the jobs where small models normally give up: fixing code, reasoning through math, following long instructions, and calling tools inside an agent workflow.

The difference is what you get for the size. OpenBMB compares it with the other open models in its weight class, Qwen3.5-2B, Gemma-4-E2B and LFM2.5-2.6B, and reports the top average score. It also edges out the 4B-class models in the same table, so you get 4B-class answers from a file half the download.

It reads about 130,000 tokens in one prompt, a few hundred pages, so a small codebase or a long report fits. Ready-made builds exist for Ollama, LM Studio, llama.cpp and Apple Silicon. The weights are Apache 2.0, so you can ship a product on it. It understands English and Chinese.

Open OpenBMB's official demo Space, paste a Python function with a bug and the error it throws, and watch it reason through the cause before handing back the fix. Then ask it to write a unit test for the repaired function. That two-step test shows exactly what this model is for. To run it on your own machine, pull the ready-made build through Ollama or LM Studio; the 4-bit versions fit comfortably in laptop memory. For tool calling, the model card points to SGLang, which turns the model's tool calls into the format agent frameworks expect.

The comparison numbers are OpenBMB's own and the model is four days old, so treat the 4B-class claim as a promise to test rather than a settled fact. A 2B model also knows far less about the world than a large one: use it where the answer can be checked, such as code and math, and expect it to invent facts when asked trivia.

• Fix a broken script offline, with no API key or usage bill.
• Run a local coding agent that calls tools such as file search or a shell.
• Work through math and logic problems inside an app that cannot send data to the cloud.
• Load a long report or a small codebase in one prompt and ask questions about it.

Try the demo

View model page

Read this issue on neuralexpedition.com →