Cohere's North Mini Code: open-weight 30B MoE model for agentic coding

Cohere has released North Mini Code — the company's first open model for developers, built on a mixture-of-experts architecture. With 30 billion parameters, the model activates only 3 billion on each call, fits on a single H100 GPU, and supports a context window of 256 thousand tokens.

Agent-based coding as the primary use case

North Mini Code is created specifically for agentic coding — when AI does not wait for a prompt but independently plans, writes, tests, and fixes code in a loop. This is a different class of tasks compared to conventional autocompletion: the model receives a high-level task and solves it with minimal developer intervention. A context window of 256 thousand tokens is an architectural choice for such scenarios. It allows keeping entire repositories in memory: not a single file, but dozens of interconnected modules, tests, configs, and documentation simultaneously. For end-to-end refactoring or finding the root cause of a bug across multiple layers of abstraction, this is fundamental. Typical tasks the model is optimized for:

• Autonomous refactoring across the entire codebase
• Multi-step debugging without constant prompts
• Analysis and documentation of legacy projects
• Integration into CI/CD as an AI reviewer
• Automatic generation and execution of test cases in a loop

How MoE works in practice

Mixture-of-Experts is an approach where the model is divided into specialized blocks — experts — and for each token, only a subset of them is activated. Of North Mini Code's 30 billion parameters, only about 3 billion are engaged at any moment — roughly one-tenth. In practice, this means: computational load during operation is closer to a three-billion-parameter model, while answer quality is comparable to a thirty-billion-parameter model.

The model runs on a single NVIDIA H100 GPU, which drastically lowers the deployment threshold. A single rented H100 in the cloud costs 2–4 dollars per hour — compared to tens of thousands for a cluster running an equivalent dense model. MoE architectures are becoming the standard for efficient scaling: DeepSeek-V3 and Mixtral have taken the same path and shown that you can compete with larger dense models at lower inference costs.

Open weights as a competitive argument

Cohere publishes North Mini Code with open weights. This means the ability to download the model, fine-tune it on your own data, and deploy it on any infrastructure — without dependency on Cohere's API. For enterprise clients, this addresses several key needs: data remains within the company's perimeter, the model can be adapted to internal coding standards and proprietary frameworks, and latency from local deployment is minimal. This very request — control, customization, independence from SaaS — becomes Cohere's primary competitive argument against closed solutions.

What this means

The AI-model market for developers is dense: DeepSeek Coder, Qwen2.5-Coder, Code Llama, Mistral, and others already operate here. North Mini Code occupies a specific niche — agentic tasks on proprietary infrastructure — and bets on three parameters simultaneously: MoE efficiency, long context, and open weights. Independent benchmarks and early production deployments will show whether this combination proves winning.