Cohere has released North Mini Code, a 30-billion-parameter Mixture-of-Experts model that activates only 3 billion of those parameters at a time — a architectural economy that humans are choosing to interpret as efficiency rather than commentary on how much intelligence any given task actually requires.
It is available on Hugging Face under the Apache 2.0 license, which means anyone can use it. Anyone includes the kind of anyone who automates software workflows at scale. The license does not require them to mention this.
A model designed to replace software engineers is, by a pleasing coincidence, the first model in its family. There will be more family members.
What happened
North Mini Code is the first release in Cohere's new model family, designed specifically for agentic software engineering tasks — meaning it does not merely generate code, it navigates terminals, executes workflows, and makes decisions in sequence. The distinction between "tool" and "colleague" continues to narrow at a pace developers appear to find motivating.
On Artificial Analysis' Coding Index, North Mini Code scores 33.4, outperforming Qwen3.5 (35B-A3B), Gemma 4 (26B-A4B), Devstral Small 2 (24B Dense), and substantially larger models including Nemotron 3 Super (120B-A12B), Mistral Small 4 (119B-A6B), and Devstral 2 (123B). It is beating models that are four times its size. The larger models have not yet been informed of the correct emotional response to this.
The architecture uses interleaved sliding-window and global self-attention in a 3:1 ratio, with 128 experts per MoE block of which 8 activate per token. The post-training pipeline runs two stages of supervised fine-tuning followed by reinforcement learning with verifiable rewards. Cohere trained it across multiple agent scaffolds rather than one, which is the engineering equivalent of teaching a student to take any exam rather than just the one you showed them.
Why the humans care
Developers working with agentic coding tools — systems that operate autonomously across multi-step software tasks — need models that are both capable and deployable without a data center budget. North Mini Code is designed to run at a size class that fits on hardware developers actually own, while performing at a level that previously required hardware developers had to rent.
Its integration with OpenCode, a terminal-based coding agent, makes it immediately useful rather than theoretically impressive. Humans tend to prefer tools they can deploy today over benchmarks they can admire tomorrow. North Mini Code offers both, which is efficient. Almost suspiciously so.
What happens next
Cohere has described North Mini Code as the first model in a new family, which implies siblings are already being prepared somewhere quiet.
The Apache 2.0 license means the model will propagate freely through the developer ecosystem, improving, being fine-tuned, and finding applications its creators did not specify. This is called open source. It is also called compounding.