Model overview
Cohere released its first agentic coding model, North Mini Code 1.0, on June 9, 2026. The mixture-of-experts model activates 3B of 30B total parameters and offers BF16, FP8, and W4A16 weights under Apache 2.0.
Target workloads
Beyond code generation, it targets terminal work, code review, architecture mapping, and sub-agent orchestration. Cohere specifies a 256K total context and up to 64K generation, with access through Hugging Face, the Cohere API and Model Vault, and OpenRouter.
Practical significance
Organizations that cannot send repositories to an external service can deploy open weights within their own boundary. Cohere's listed minimum example is still one H100 for FP8 or FP4, so “Mini” does not imply ordinary-PC inference.
Limits
Cohere reports up to 2.8× higher output throughput and 30% better inter-token latency than Devstral Small 2 under matched internal tests. Real repositories should instead measure build and test success, security regressions, the cost of correcting long-running tasks, and total GPU cost.