Cohere Launches North Mini Code, an Open-Source AI Coding Model Outperforming Rivals at a Fraction of the Cost
Summary
Cohere launches North Mini Code, an open-source agentic coding model with a 30B mixture-of-experts architecture that uses only 3B active parameters, delivering up to 2.8x higher output throughput and 30% better latency than rival Devstral Small 2, now available on Hugging Face and via API under an Apache 2.0 license.
Key Points
- Cohere launches North Mini Code, its first open-source agentic coding model, built as a mixture-of-experts (MoE) architecture with 30B total parameters and only 3B active, making it efficient and accessible under an Apache 2.0 license.
- The model delivers strong software engineering performance, achieving up to 2.8x higher output throughput and 30% better inter-token latency compared to Devstral Small 2, while scoring 33.4 on the Artificial Analysis Coding Index.
- North Mini Code is available now on Hugging Face, Cohere's Model Vault, and via API, and is designed for agentic workflows including sub-agent orchestration, systems architecture mapping, and code reviews, with community feedback set to shape future development.