On June 9th, Cohere released North Mini Code under the Apache 2.0 license, with weights on Hugging Face. It’s the company’s first model aimed squarely at developers — and the pitch isn’t “we beat GPT”. It’s something more interesting for anyone who’s ever hesitated before pushing a private repo to a hosted API: a capable coding agent, small enough to run on hardware you control.
What it is, specifically
North Mini Code is a Mixture-of-Experts model: 30B total parameters, but only 3B active per token. That ratio is the whole story. You get the representational capacity of a medium-sized model with an inference cost closer to a small one — which is exactly what makes self-hosting realistic instead of aspirational.
Cohere built it specifically for agentic software engineering, not general chat: orchestrating sub-agents, mapping unknown architecture, doing code review, and working in the terminal. It’s the first model in what Cohere calls a new family, and the framing of the entire launch is “sovereign” development — keeping your code on your own infrastructure instead of sending it to someone else’s endpoint.
The numbers, honest framing
Cohere reports 80.2% pass@10 on SWE-Bench Verified for the fine-tuned model, and 61.0% pass@1 across the mini-SWE-Agent harness. These are solid numbers for a model of this size.
The headline claim — that it beats open-source rivals several times its size — is real but worth clarifying. It comes from the Coding Index by Artificial Analysis, where North Mini Code scores 33.4, ahead of models like Devstral 2 (123B) and Nemotron 3 Super (120B-A12B). So the “punches above its weight” story is specific to that benchmark, not a claim across all evaluations. It’s still a strong result: a 30B model outperforming things five times larger on a respected third-party index is the kind of efficiency that changes the math of deployment.
Something Cohere highlights and that’s easy to miss: they trained it across multiple agentic harnesses instead of optimizing for a single one. That matters in practice — a model that only performs well within the exact scaffold it was tuned for tends to fall apart the moment you change the tooling. North Mini Code is built to hold up across SWE-Agent, mini-SWE-Agent, and OpenCode, which is the difference between a benchmark trophy and something you can actually build an agent on.
Why your team should care
The real hook isn’t the leaderboard position. It’s that an agentic, open-weight, Apache 2.0 coding agent of this size makes a concrete option viable: running your agent on your own hardware, where your source code never leaves the building.
For many teams that’s not a luxury. Regulated industries, work under NDA for clients, internal codebases that legal would prefer not to see flowing through a third-party API — everyone has lived with a real tension between wanting agentic tooling and not being allowed to use hosted versions. A model you can download and serve yourself dissolves that tension, at least in principle.
The cost angle is also part of it. Cohere is explicit about aiming for a lower total cost of ownership, and the MoE design backs it up: you’re not paying to activate 30B parameters on every token. For teams running agents at volume, the difference between an API bill per token and self-hosted inference on your own hardware adds up fast.
Where to get it
The weights are on Hugging Face under Apache 2.0. If you’d prefer not to run it yourself, Cohere also offers it via their API, through Model Vault (their dedicated, managed inference option), and you can try it in harnesses like OpenCode.
The honest caveat: self-hosting a model is never just “download and go” — you need the infrastructure and the willingness to operate it. But for the first time, the model itself isn’t the bottleneck. That’s the change worth paying attention to.