Cohere released North Mini Code, an open-weight coding model built to run on a single Nvidia H100 GPU, its first model aimed squarely at developers rather than enterprise search. It is a roughly 30-billion-parameter mixture-of-experts agentic coder with a 256K-token context, published as open weights and free to use. The pitch is unusual in a field obsessed with frontier scores: a model you can self-host on one GPU, keep your source code inside your own network, and still get real agentic coding out of. For any team that cannot ship its codebase to a hosted API, that combination matters more than a benchmark crown.
- North Mini Code is an open-weight ~30B mixture-of-experts agentic coding model, Cohere's first developer-focused release.
- It runs on a single H100 GPU with a 256K-token context and is publicly available and free to use.
- The target is self-hosted, private deployment: coding help for banks, governments and regulated firms that will not send source to a cloud API.
- It fits Cohere's enterprise, data-sovereignty niche rather than chasing the frontier score race that Fable 5, GPT-5.6 and GLM-5.2 are fighting.
What is North Mini Code and why is it different?
North Mini Code is Cohere's first model aimed at developers rather than the enterprise search and retrieval work the company is known for. It is a mixture-of-experts design of roughly 30 billion parameters, which is the key to its headline feature: an MoE only activates a fraction of its parameters per token, so the model punches above its footprint and fits on a single H100 instead of demanding a cluster. Cohere published it as open weights with a 256K-token context and made it free to use. In a landscape where the loudest launches are gated frontier models, a self-hostable coder that runs on one GPU is a deliberately different bet.
RelatedGoogle Ships Gemini 3 Pro Image and 3.1 Flash Image
Why does single-GPU, self-hosted matter?
The constraint is the point. A model that fits on one H100 can run inside a company's own data center, which means a developer's source code, prompts and outputs never leave the corporate network. For most consumer developers that is a nice-to-have. For a bank, a hospital, a defense contractor or a government agency, it is often a hard requirement, because sending proprietary or regulated code to a third-party cloud API is either against policy or against the law. Cohere has always aimed at exactly these customers, and North Mini Code extends that data-sovereignty pitch from search into coding, where the sensitivity of the input is even higher.
Self-hosted open coder versus hosted frontier
| Trait | North Mini Code | Hosted frontier coder |
|---|---|---|
| Weights | Open, free | Closed, API-only |
| Hardware | Single H100 | Multi-GPU cluster |
| Your code | Stays on-prem | Sent to provider |
| Peak capability | Good, not frontier | Highest available |
How does mixture-of-experts make this possible?
The engineering that lets a 30B model behave like something larger while running on modest hardware is the mixture-of-experts architecture. Instead of routing every token through the full network, an MoE has many specialized subnetworks and activates only the few most relevant ones per token. The result is a model with the knowledge capacity of its full parameter count but the runtime cost of the smaller active slice. That is why North Mini Code can target a single H100: the memory footprint is manageable and the per-token compute stays low. MoE is the same trick powering the biggest open models of 2026, applied here to shrink a capable coder down to one accelerator.
Where does this leave Cohere in the model race?
Cohere is not trying to win the frontier benchmark war, and North Mini Code makes that clear. While Anthropic, OpenAI and the Chinese open-weight labs trade the top of the coding leaderboards, Cohere is carving out the enterprise flank: models that are private, deployable and cheap to run, sold to organizations that value control over raw capability. A free, open-weight coder that runs on one GPU is a smart wedge into developer workflows at exactly those accounts. It will not top a SWE-bench chart, but topping charts was never the plan. Owning the regulated, on-prem coding market is a smaller prize that is far harder for a hosted-only rival to contest.
RelatedZhipu's GLM-5.2 Tops the Open-Weight Model Rankings
- Independent coding scores. Whether North Mini Code posts a credible SWE-bench or Terminal-Bench number to anchor its quality claim.
- Enterprise pull. Named deployments at banks or government agencies would prove the private-coding thesis.
- Cohere's cadence. Whether this is a one-off or the start of a full developer-model line to rival hosted coders.
Our take
North Mini Code is a refreshing counter-move in a month dominated by gated frontier launches: instead of a bigger model behind a stricter gate, Cohere shipped a capable one you can actually run yourself. The single-H100 constraint plus open weights plus a genuine data-sovereignty story is a coherent product, not a benchmark stunt, and it targets customers the frontier labs cannot easily serve without an on-prem offering. The open question is quality. Without a published independent coding score, "good enough to self-host" is a claim, not a fact, and enterprises will demand proof. If the numbers hold up, this is the most strategically sound coding release of the month, precisely because it is not chasing the leaderboard.
- OfficialCohere North Mini Code and North platform
- BenchmarkGenZTech AI Coding Leaderboard verified coding-model standings
- ReferencePrice Per Token, model releases recent open-weight launches
Original analysis by GenZTech. Reporting informed by LLM-Stats.