Cohere launches North Mini Code, its first open-source agentic coding model

Cohere has released North Mini Code open-source, marking the company's first agentic coding model and the first entry in what it describes as its next generation of powerful models. The model uses a mixture-of-experts (MoE) architecture with 30B total parameters and just 3B active parameters, a design intended to deliver strong software development performance while minimizing hardware requirements. The minimum hardware requirement listed is a single H100 GPU at FP8 precision.

Cohere frames the release as part of its broader mission to make sovereign AI practical for developers.

The model is released under an Apache 2.0 license and is available in several ways: weights can be downloaded from Hugging Face, it can be deployed in a managed inference environment via Cohere's Model Vault, accessed through the Cohere API, or tried via OpenRouter. North Mini Code supports a 256K total context length with a 64K maximum generation length, and is optimized for code generation, agentic software engineering, and terminal tasks. Cohere frames the release as part of its broader mission to make sovereign AI practical for developers.

Key facts

1. 01North Mini Code is Cohere's first agentic coding model and first open-source model for developers.
2. 02It uses a mixture-of-experts (MoE) architecture with 30B total parameters and only 3B active parameters.
3. 03Released under an Apache 2.0 license.
4. 04Supports a 256K total context window with a 64K maximum generation length.
5. 05Optimized for code generation, agentic software engineering, and terminal tasks.
6. 06Minimum hardware requirement is 1× H100 GPU at FP8 precision.
7. 07Available via Hugging Face (weights), Cohere API, Cohere Model Vault, and OpenRouter.

Topics