Cohere opens North Mini Code
On 9 June 2026 Cohere released North Mini Code, a mixture-of-experts model with 30 billion parameters and 3 billion active, built for agentic programming, and published its weights on Hugging Face under the Apache 2.0 licence.
Why it matters
A model trained specifically for agentic work on code and runnable on one’s own hardware became available under a free licence. An editorial assessment.
By Cohere's description the model is a decoder with 128 experts, 8 active per token, interleaving sliding-window and global attention in a ratio of 3 to 1. Post-training: two stages of supervised fine-tuning and reinforcement learning with verifiable rewards on more than 70 thousand tasks from about 5 thousand repositories, deduplicated against the SWE-Bench and SWE-Bench-Pro repositories. The weights come in BF16 and FP8; the model is also in OpenCode and through the Cohere API. The repository CohereLabs/North-Mini-Code-1.0 carries the Apache 2.0 licence. What the record does not claim. Cohere reports a score of 33.4 on the Artificial Analysis Coding Index itself; we did not check it and so give no comparison with other models. The post names no hardware requirement. 'First model' refers to the new North family, not to all of the company's open weights.