North Mini Code

North Mini Code is Cohere's purpose-built open-weight checkpoint for code generation, agentic software engineering, and terminal tasks. The 30-billion-parameter MoE activates 3 billion parameters and was trained with tool use for coding agents.

Cohere documents 256,000 context tokens and 64,000 output tokens. The model card requires a suitable chat template for reliable tool calls and recommends carrying thinking and tool content through the conversation. Apache 2.0 permits self-hosting. For vLLM, Cohere currently calls for a main development build and a dedicated parser.