North Mini Code
A 30B sparse mixture-of-experts model for code and terminal agent work
North Mini Code is a downloadable research model built for writing code, driving software agents and handling terminal work. Cohere and Cohere Labs released it as open weights, with 30 billion parameters in total but only 3 billion used per token, plus a 256K-token context window.
- 30 billion parameters in total, 3 billion active for each token
- Context window of 256K tokens, with up to 64K tokens of output
- 128 experts in the feed-forward block, 8 used per token
- Post-trained with supervised fine-tuning then reinforcement learning on verifiable rewards
- Weights published for download, with a hosted Space and OpenCode for trying it
What it is
North Mini Code is a research release of a language model aimed at code generation, agentic software engineering and tasks run in a terminal. Cohere and Cohere Labs put the weights out for anyone to download. The company describes it as a 30B-A3B model: 30 billion parameters overall, of which 3 billion are used for each token it produces. It accepts 256K tokens of context and can write up to 64K tokens in one reply. It outputs text, and it has been trained to call tools so it can act inside a software project rather than only answer questions.
How it came about
Cohere and Cohere Labs developed the model and published it under the name North-Mini-Code-1.0 on Hugging Face, with a blog post giving further technical detail. The release is framed as research, and the point of contact for questions is Cohere Labs. No earlier version of this model is described, and no release date is given.
Inside the model
The design is a decoder-only Transformer built as a sparse mixture of experts, meaning most of the network stays idle for any one token. Attention alternates between a sliding-window form using RoPE position encoding and a global form with no position encoding, in a three-to-one pattern. The feed-forward part holds 128 experts and turns on 8 of them per token; each expert is an FFN with SwiGLU activation, and the router applies a sigmoid to its logits before picking the top few. One dense layer sits ahead of the sparse ones. Training finished with two rounds of supervised fine-tuning, then reinforcement learning with verifiable rewards, aimed at agentic coding.
What it can do
The company measured it on SWE-Bench Verified, SWE-Bench Pro, Terminal-Bench v2 and Terminal-Bench Hard for agent work, plus SciCode and LiveCodeBench v6 for code writing without tools. Runs used three seeds and were averaged, at temperature 1.0 and top_p 0.95. Competitor figures came from public reports or the Artificial Analysis Intelligence Index, and some missing numbers were produced internally. No scores are given here. The model supports tool calls through chat templates, works with interleaved thinking, and the company says it performs best with that switched on.
How to use it
Weights are downloadable from Hugging Face. Before fetching them you can try the model in OpenCode or in a hosted Hugging Face Space. It runs through Transformers, vLLM and OpenCode; vLLM needs the main branch plus Cohere's melody library for correct parsing of replies, and OpenCode likewise needs its main branch.
Safety and testing
The model card describes training aimed at agentic coding and tool use, but sets out no safety evaluations, red-teaming or incidents.
Reception and debate
No reviews, criticism, legal actions or bans are described.
Why it matters
The company places the model among its other agent models, all of which share support for interleaved thinking. Nothing further is said about uptake or models built on top of it.
Questions people ask about North Mini Code
What is North Mini Code for?
Cohere built it to write code, carry out software engineering tasks as an agent, and work inside a terminal. It can call tools such as a bash command, and it is meant to be run through chat templates that pass tool descriptions in JSON schema form.
How big is it and how much of it runs at once?
The network holds 30 billion parameters, but only 3 billion of them are used for any single token. That is what the 30B-A3B label means. It handles 256K tokens of input and can generate up to 64K tokens in one go.
How do I run it?
Download the weights from Hugging Face and load them with Transformers, which must be built from the source repository. vLLM and OpenCode also work, though both need their main branch for now, and vLLM additionally needs Cohere's melody library. A hosted Space lets you test it first.
What kind of licence does it carry?
The company calls it an open weights research release and lets you download the weights, but the licence terms themselves are not stated in the model card.
How was it evaluated?
Cohere ran it on four agentic coding benchmarks and two code generation benchmarks, averaging three seeds per run at temperature 1.0 and top_p 0.95. Competitor numbers were taken from public reports or the Artificial Analysis Intelligence Index where those existed.
Good reading and sources
There is no English Wikipedia article on North Mini Code yet, so the facts on this page come from the sources listed above: huggingface.co/CohereLabs/North-Mini-Code-1.0. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.