Cohere and Cohere Labs have released North Mini Code, a 30B-parameter mixture-of-experts coding model with 3B active parameters, Apache 2.0 weights, a 256K-token context window, and a 64K-token maximum output length. Hugging Face source documents position the model for code generation, agentic software engineering,...
Cohere and Cohere Labs released North Mini Code, an open-weights coding model aimed at developer and software-engineering tasks.
In a Hugging Face blog post titled “Introducing North Mini Code: Cohere’s First Model For Developers,” CohereLabs describes North Mini Code as a 30B-parameter mixture-of-experts model with 3B active parameters. The accompanying Hugging Face model card for CohereLabs/North-Mini-Code-1.0 similarly describes it as a 30B-A3B model optimized for code generation, agentic software engineering, and terminal tasks.
The release uses Apache 2.0 licensing for the weights, according to the Hugging Face blog and model card. That license choice is notable for developers and organizations evaluating whether they can inspect, modify, or deploy the model weights in their own environments, subject to the license terms.
CohereLabs says the model supports a 256K-token context window and a maximum output length of 64K tokens. Those limits position North Mini Code for tasks involving larger codebases or longer generated outputs, although real-world usefulness will depend on deployment constraints, inference cost, and task quality.
The Hugging Face blog says Cohere evaluated North Mini Code with methodology covering SWE-Bench, Terminal-Bench, SciCode, and LiveCodeBench. These benchmark references indicate the release is being framed around practical software-engineering work rather than only short code-completion prompts.
The CohereLabs/North-Mini-Code-1.0 model card states that the model is optimized for code generation, agentic software engineering, and terminal tasks. The blog’s benchmark set aligns with that positioning: SWE-Bench is commonly used to assess issue-resolution behavior in software repositories, while Terminal-Bench focuses on terminal-oriented task execution. SciCode and LiveCodeBench are also cited by CohereLabs as part of the model’s evaluation coverage.
The available source excerpts do not provide enough detail to independently compare North Mini Code’s scores against other coding models or verify performance claims beyond CohereLabs’ stated benchmark methodology. For that reason, the release is best read as an availability and technical-specification announcement rather than a definitive ranking of coding-model capability.
CohereLabs also published CohereLabs/North-Mini-Code-1.0-fp8 on Hugging Face. The FP8 model card describes the quantized North Mini Code weights and repeats the same core positioning: 30B total parameters, 3B active parameters, Apache 2.0 licensing, and a focus on coding-agent use cases.
The availability of an FP8 variant may matter to developers testing deployment options, because quantized weights can reduce memory requirements compared with higher-precision formats. The provided source excerpt does not specify performance trade-offs for the FP8 version, so users would need to consult the model card and run their own evaluations for target workloads.
For teams considering North Mini Code, the Hugging Face blog and model cards provide the starting point: model size, active-parameter count, license, context length, output length, and intended use cases. Developers should also review the full model documentation for supported inference setups, safety notes, evaluation details, and any usage recommendations that are not included in the short release excerpts.
Because the model is presented as open weights under Apache 2.0, it may appeal to teams that want more control than a hosted-only coding assistant can provide. At the same time, coding models can produce incorrect patches, insecure code, or terminal commands that need review. CohereLabs’ framing around agentic software engineering makes careful sandboxing, testing, and human oversight especially important when connecting the model to repositories, shells, or deployment systems.
North Mini Code adds a new developer-focused model to Cohere’s public model lineup, with the main technical claims documented by CohereLabs on Hugging Face: a 30B-A3B mixture-of-experts architecture, Apache 2.0 weights, long-context support, and benchmark coverage across several code and software-engineering evaluations.
Cohere and Cohere Labs released North Mini Code, an open weights coding model aimed at developer and software engineering tasks.
The accompanying Hugging Face model card for CohereLabs/North Mini Code 1.0 similarly describes it as a 30B A3B model optimized for code generation, agentic software engineering, and terminal tasks.
The release uses Apache 2.0 licensing for the weights, according to the Hugging Face blog and model card.
Continue reading
