Cohere released North Mini Code on June 9, offering a downloadable model designed for coding agents and terminal-based development tasks. The launch announcement lists an Apache 2.0 license, access through hosted services and a model architecture with 30 billion total parameters, of which about three billion are active during computation.
That mixture-of-experts design selects parts of the model for each step. The active-parameter figure describes the work involved in processing a token; it does not mean the complete model can be stored as if it had only three billion parameters.
Developers can choose weights or hosted access
Cohere lists downloadable weights on Hugging Face, its API, Model Vault and OpenRouter among the access routes. It also describes compatibility with OpenCode and other coding-agent software.
The published specifications include a total context window of 256K tokens and a maximum generation length of 64K tokens. These are capacity limits, not a guarantee that every long codebase will be understood correctly or that a task will need the full allowance.
Downloadable weights give teams a route to controlling where inference runs. That choice brings a deployment requirement: the model, its context and the serving software need memory and computing capacity. The name “Mini” is relative to other models, not evidence that any laptop can run it comfortably.
Our guide to running AI locally explains the difference between obtaining model files and operating a usable local service.
Coding-agent results depend on the surrounding software
Cohere reports coding benchmarks and compares generation performance with Devstral Small 2. The company says North Mini Code reached up to 2.8 times higher output throughput in its tests, while the comparison model retained a slight advantage in time to first token under the tested conditions.
That is a more specific result than saying every coding task finishes nearly three times faster. An agent’s completion time also includes reading files, calling tools, running tests and recovering from unsuccessful attempts.
The release notes the harnesses used for different evaluations, including SWE-agent and terminal-oriented setups. Teams comparing models can preserve those distinctions when evaluating their own repositories. A benchmark score is most useful when the tools and task conditions behind it are understood.
