Last updated: 2026-07-23
Last verified: 2026-07-23
TL;DR: North Mini Code is Cohere Labs’ open-weight coding model for agentic software-engineering and terminal tasks. It has 30 billion total parameters, activates about 3 billion per token through a mixture-of-experts design, supports a 256K context window, and is released under Apache 2.0. Its compact active footprint is attractive, but deployment and benchmark claims still need local validation.
What Cohere Labs released
Cohere Labs introduced North Mini Code on June 9, 2026 as the first model in a new North family. The release targets coding-agent workflows rather than only autocomplete or isolated function generation. Its public materials emphasize repository work, terminal interaction, tool use, and longer sequences of edits and tests.
The model uses a mixture-of-experts architecture. Although the checkpoint contains 30 billion parameters, roughly 3 billion are active for each token. That design aims to offer more model capacity than a dense model with a similar active compute requirement, but memory use still reflects the full weights and the selected quantization and serving setup.
Model specifications
The official model card lists a 256K context window and a maximum output length of 64K tokens. It provides examples for Transformers and points to vLLM and SGLang for serving. Cohere Labs also describes use with OpenCode and a hosted Space for initial exploration.
- 30B total parameters with approximately 3B active parameters.
- Open weights under the Apache 2.0 license.
- Long context intended for repository-scale inputs and tool history.
- Support paths for common open-model serving frameworks.
- Training focused on agentic coding and terminal interaction.
Where it may fit
The model is a candidate for teams that want more control over coding-agent inference, including the ability to run inside their own environment, inspect prompts, and choose a serving stack. It may suit repository triage, bounded bug fixes, test generation, refactoring proposals, and shell tasks where a reviewer can inspect every change.
Open weights do not make the full system local by default. Hosted demos, third-party inference providers, telemetry, package downloads, and agent tools can still send data elsewhere. A deployment review should trace the complete path from repository content to model server, logs, and external tools.
How to evaluate the model
Cohere Labs publishes benchmark results from several agent evaluation systems. Those numbers are useful for choosing models to test, but they do not predict performance on a private repository. Evaluation configuration, tool availability, retry budgets, and test quality can materially change the result.
- Select ten closed issues with known fixes and a mix of languages and repository sizes.
- Give each model the same tool permissions, context, token budget, and time limit.
- Measure correct patches, regressions, unnecessary edits, test results, and reviewer time.
- Record GPU memory, tokens per second, queue time, and total infrastructure cost.
- Inspect license notices and dependency obligations before distributing a derived system.
Deployment considerations
A 256K context limit does not mean every task should fill it. Large prompts increase latency and can bury the files that matter. Retrieval, repository maps, and explicit task boundaries may produce better results than sending an entire codebase. Long output limits also require controls so the agent does not spend its budget narrating or making broad changes.
For production use, isolate the execution environment, restrict network access, avoid exposing secrets, and make commits reviewable. The model can propose terminal commands, but an orchestration layer should decide which commands are permitted and when a person must approve them.
Pricing and license
The weights are downloadable under Apache 2.0, so there is no mandatory per-token fee from Cohere Labs for self-hosting. Compute, storage, bandwidth, engineering, monitoring, and support remain real costs. A hosted provider may charge separately and impose its own data and availability terms.
Review the model card and license files for the exact checkpoint used. “Open weight” describes access to the parameters; it does not certify a model for a regulated workflow, remove third-party code obligations, or guarantee that generated code is safe to ship.
Kingy AI verdict
North Mini Code is worth benchmarking for teams that want a relatively low active-parameter coding model with long context and open deployment options. The strongest case is controlled repository work with repeatable tests. Compare it with both smaller local models and managed coding agents on the same issue set before choosing a serving stack.
FAQ
How large is North Mini Code?
It has 30 billion total parameters and activates about 3 billion parameters per token.
Can it be self-hosted?
Yes. The model card provides open weights and points to Transformers, vLLM, and SGLang deployment paths.
Is it free to run?
The weights do not carry a per-token Cohere fee for self-hosting, but hardware, operations, and review costs still apply.
Official links
Related Kingy AI links
Kingy Launch Brief
Put the week’s verified AI launches in your inbox.
Every Friday, the verified AI launches, apps, funding rounds, pricing changes and under-the-radar moves worth knowing—source-linked and explained in five minutes.
Free · Every Friday · Unsubscribe anytime · No daily email