Cohere's North Mini Code is a 30B open MoE built for coding agents

Cohere Labs released North Mini Code, the company's first developer-focused model, under Apache 2.0. It's a 30B-parameter mixture of experts with 3B active per token, sized for coding agents rather than chat. On the Artificial Analysis Coding Index it scores 33.4, ahead of much larger models including Nvidia's Nemotron 3 Super at 120B and Mistral Small at 119B. On SWE-Bench Verified it hits 80.2% pass@10 after supervised fine-tuning, and on Terminal-Bench v2 it lands at 55.1% pass@10, with an extra 7.9 absolute points on pass@1 after reinforcement learning with verifiable rewards.

The training story is the unusual part. Cohere ran two stages of SFT followed by RLVR on more than 70,000 containerized agentic tasks drawn from roughly 5,000 repositories. They also trained the model on multiple agent scaffolds in parallel, including SWE-Agent, mini-SWE-Agent, and OpenCode, so it isn't tuned to a single harness. The context window during RL was 128K. The result is a small-active-parameter model that runs cheaply in an agent loop while staying robust to whichever loop you build around it.

Why it matters

If you ship a coding agent, the question this raises is whether you can drop in a 3B-active open model for your inner loop and keep the bigger closed models only for the hard cases. North Mini Code's numbers suggest the answer is yes for a wide slice of tasks, and the multi-scaffold training reduces lock-in to whichever harness you started with. Apache 2.0 means you can fine-tune and self-host without legal review.