30B total / 3B active (MoE)
256K
Apache 2.0
Official open-weight coding-focused release positioned by Cohere as its first model for developers and the first member of a North family of code agent models. Primary release materials describe North Mini Code as a decoder-only sparse Mixture-of-Experts model optimized for code generation, agentic software engineering, and terminal tasks, with 30B total parameters and 3B active parameters. The Hugging Face model card says the architecture uses interleaved sliding-window RoPE attention and global attention in a 3:1 ratio, 128 experts with 8 activated per token, SwiGLU feed-forward experts, and a router that applies sigmoid before top-k selection, plus a single dense layer before the sparse stack. Cohere states the model was post-trained with a two-stage cascaded supervised fine-tuning pipeline followed by reinforcement learning with verifiable rewards focused on agentic coding. Release docs publish a 256K context window with up to 64K output tokens and position the model for repo-level code changes, terminal-based agents, local deployment, and general code-generation tasks. The launch post emphasizes competitive benchmark performance for its size on software-engineering and terminal-agent evaluations, including comparisons against similar open models, and claims up to 2.8x higher output throughput than Devstral Small 2 in Cohere's internal testing under matched hardware and concurrency settings. Distribution is provided through Hugging Face weights, Cohere API, Model Vault, and OpenCode compatibility guidance.