North Mini Code: Open-Weight Coding Model, Specs, and Evaluation

This older record retains its original scope. Current buying qualifications above take precedence where terms have changed.

Kingy verdict: North Mini Code is notable because it combines open Apache 2.0 weights, a 30-billion-parameter mixture-of-experts architecture with 3 billion active parameters, and explicit training for agentic software engineering. The small active count can improve inference economics, but it does not reduce the stored model to 3 billion parameters or remove the need to test a complete serving stack.

What North Mini Code is

Cohere released North Mini Code on June 9, 2026 as the first model in its North family. The official model card lists 30 billion total parameters, 3 billion active parameters per token, Apache 2.0 licensing and a text-focused coding-agent design. Cohere’s documentation lists a 256K-token context window, up to 64K output tokens and the API identifier north-mini-code-1-0.

The model is designed for repository-level changes, terminal agents and software-engineering harnesses. Cohere says its training spans multiple agent scaffolds rather than one fixed interface. That is useful positioning, but it does not guarantee identical performance across an organization’s tools, prompts, containers or permission model.

Access and cost

The weights are available through the official Cohere Labs Hugging Face repository. Cohere’s current documentation also says trial and production API keys can use the model free until applicable rate limits are reached, while production deployment is available through Model Vault. “Free” therefore describes the documented API meter or the downloadable license, not hardware, engineering, support, observability or security cost.

Local buyers should size memory against the full 30-billion-parameter weight set, chosen precision, KV cache, context length and serving concurrency. Three billion active parameters describes routing during inference; it is not a complete capacity plan.

How to evaluate it

Run the same repository task in at least two harnesses. Require the model to inspect the codebase, propose a plan, edit multiple files, run tests, recover from one failed command and provide a reviewable diff. Measure task completion, valid tool calls, test pass rate, unnecessary churn, time to first accepted patch, total generated tokens and peak serving memory. Add an adversarial case that checks whether the agent respects file and command boundaries.

Cohere and Cohere Labs publish benchmark results and methodology notes, including cautions about harness differences. Kingy did not reproduce those evaluations. The defensible comparison is therefore a fixed local task pack under the buyer’s actual harness and hardware, with provider scores retained as context rather than treated as an independent verdict.

Primary sources

• Cohere launch announcement
• Cohere model documentation
• Official Cohere Labs model card and weights
• Cohere Labs technical introduction and benchmark methodology