North Mini Code opens Cohere's North family as a coding-focused model designed for agentic software engineering rather than general conversation. It uses a sparse mixture-of-experts design with 30B total parameters but only 3B active per token, a footprint that keeps inference lightweight enough for local hardware while still leaving room to handle large codebases. The model is tuned to drive tool-using developer agents, support sub-agent orchestration, multi-file analysis, code review, and real terminal workflows, and it is trained to generalize across common agent harnesses such as OpenCode and SWE-Agent. It also supports interleaved reasoning and structured tool calls via JSON schema, so it can plan a change, call tools, and continue reasoning in a single loop without hand-holding from a wrapper framework.

As Cohere's first release in the new North lineup, North Mini Code signals a clear direction toward compact, open-weight specialists aimed at developer workflows. The sparse active-parameter design, the agent-harness generalization during training, and the Apache 2.0 release together position it as a base developers can fine-tune, distill, or drop into existing pipelines rather than a closed assistant. The mix of open weights, structured tool use, and coding-specific post-training makes it a strong fit for teams building IDE copilots, repository-scale refactoring agents, terminal assistants, and multi-step coding loops where responsiveness and controllability matter as much as raw capability.