Cohere Presents North Mini Code — Model for Developers and AI Agents

Cohere released North Mini Code — its first model specifically developed for programmers. The 30-billion-parameter model with 3 billion active parameters is now available on Hugging Face under the Apache 2.0 license, trained on programming, AI agent work, and terminal interaction.

Architecture and Performance

North Mini Code is built on a Mixture of Experts architecture: 128 experts, of which 8 are active at any given time. The model uses an attention mechanism in a 3:1 ratio (local and global), as well as SwiGLU activation. On coding quality, it scores 33.4 points on Artificial Analysis' Coding Index — outperforming Qwen3.5 (35B), Gemma 4 (26B), and even Mistral Small (119B), despite those being larger in size.

Two-Stage Reinforcement Learning Training

At the first stage of supervised fine-tuning, Cohere focused on programming: 70% of data is code, 43% are examples of tool interaction (agentic tool-use), 27% are solutions to competitive and scientific problems. The second stage used 4.5 billion tokens of the highest quality to refine behavior in agent scenarios. Then they applied reinforcement learning with verifiable rewards (RLVR), training the model on two types of tasks simultaneously:

• Terminal work (bash commands, output and error processing)
• Code editing and debugging in files
• Tool invocation via API and standard input/output

Trained on Multiple Frameworks Simultaneously

Cohere trained North Mini Code not on a single agent framework, but on three at once — SWE-Agent, mini-SWE-Agent, and OpenCode. This is critical: each framework requires its own command format and result processing. This approach makes the model universal and reliable in any real-world environment.

"We trained on multiple agent frameworks instead of optimizing for a single interface to make the model robust to different tools,"

Cohere explained.

Results: 7-8% Better

After RLVR training, the model improved results by 7.9% on Terminal-Bench v2 and by 3.0% on SWE-Bench (compared to the base version). Human evaluation on 85 examples showed 66.1% wins on code editing tasks. The model makes significantly fewer errors when working with tools and handles long command sequences more confidently.

What This Means

North Mini Code is a signal that developers no longer need general-purpose LLM models. Companies are training models specifically for code and agents because it works better. For developers, this means: you can deploy a powerful AI assistant locally without dependence on cloud APIs and paid services.