Cohere launches North Mini Code, an open-weight code model with 320K context

Cohere officially launched North Mini Code, a compact yet powerful code generation model with open weights. The model supports a 320K token context window and is available for download on Hugging Face in fp8 format. Developers can also try it for free on the OpenCode platform. For self-hosting, Cohere recommends using vLLM main branch (until a new release) with tensor parallelism (e.g., -tp 2) and installing the cohere_melody library for accurate response parsing. The model includes built-in tool-call and reasoning parsers, making it suitable for agentic code workflows.

Beyond deployment, North Mini Code has garnered early community support: a third-party MLX conversion enables efficient inference on Apple Silicon. Cohere acknowledges demand for quantization and llama.cpp support, signaling future updates. The launch reflects Cohere's commitment to developer-friendly, open models for code, with active feedback loops to refine capabilities. For teams building code assistants or automated development pipelines, this model offers a lightweight, long-context alternative to larger proprietary systems.