Cohere just open-sourced its first agentic coding model, North Mini Code, the debut of a new line of efficient models. It uses a Mixture of Experts (MoE) architecture with 30 billion total parameters but only activates 3B per forward pass. Under an Apache 2.0 license, it can run locally or in private clouds — giving developers what Cohere calls "sovereign AI," or independence from cloud vendors.
In Artificial Analysis's Coding Index, North Mini Code scored 33.4, competitive for its parameter class. On identical hardware and concurrency, it delivered 2.8 times the output throughput of Devstral Small 2, meaning faster iteration and lower compute costs. Its inter-token latency — a measure of generation consistency — is 30% lower than Devstral Small 2, and time to first token (TTFT) is roughly the same.
North Mini Code was built specifically for agent workflows. It's good at coordinating sub-agents, drawing system architecture diagrams, and reviewing code. You can grab the weights on Hugging Face as CohereLabs/North-Mini-Code-1.0, or try it through Cohere's managed inference platform Model Vault, OpenCode, and the official API for free.