Cohere launches North Mini Code 1.0, its first agentic coding model, as open weights under Apache 2.0

Cohere released North Mini Code 1.0 on June 9, 2026 — its first model purpose-built for agentic coding workflows. The model is a 30-billion parameter Mixture of Experts (MoE) architecture with only 3 billion active parameters per forward pass, making it lightweight enough to run on local hardware while maintaining a 256K input / 64K output token context window.

What's new

North Mini Code 1.0 (north-mini-code-1-0) ships with:

• Architecture: 30B total parameters, 3B active (MoE), enabling local deployment
• Context window: 256K input tokens, 64K output tokens
• License: Apache 2.0 (fully permissive for commercial use)
• Access: Available via the Cohere Chat V2 API and as open weights on Hugging Face; Model Vault deployment is also supported for production use

The model is positioned as "Cohere's first agentic coding model," optimized for multi-step, tool-using coding tasks rather than simple code completion. The MoE design means that despite the large total parameter count, inference cost and latency are comparable to a 3B dense model — making it practical for production agentic pipelines that might invoke the model repeatedly per task.

Context

Cohere has historically focused on enterprise text embedding, reranking, and command models rather than code-specific products. North Mini Code represents the company's entry into the coding model market, which has become increasingly competitive with dedicated releases from Anthropic (Claude's coding capabilities), Microsoft (MAI-Code-1), and open-weight options from Qwen, DeepSeek, and others.

The Apache 2.0 license and open weights on Hugging Face are notable choices for Cohere, a company that has traditionally kept its frontier models API-only. Releasing open weights positions North Mini Code as a credible option for enterprises that need to run coding agents in air-gapped or on-premise environments where API calls to external providers are not permitted.

The MoE architecture — 30B total, 3B active — follows the pattern established by DeepSeek-MoE, Mistral's Mixtral, and Qwen-MoE models, where a large total parameter pool (providing capacity and knowledge diversity) is paired with sparse activation (keeping inference cost down). For agentic coding specifically, where the model may be called dozens of times per task, this cost profile is meaningful.

Why it matters

North Mini Code fills a gap in the open-weights coding model landscape: a model explicitly designed for agentic workflows, released under Apache 2.0, and engineered to run locally. Most existing open-weights coding models (CodeLlama, StarCoder, DeepSeek-Coder) were designed primarily for single-turn code completion rather than multi-step agent loops with tool calls and extended context.

For enterprises building internal coding agents — code review, automated testing, refactoring pipelines — a 3B-active-parameter model running locally can keep inference costs near zero while avoiding data exposure to external APIs. The 256K context window is also large enough to handle real codebases, where relevant files often span tens of thousands of tokens.