Cohere North Mini Code Gives Agent Builders a 30B Open Coding Model
Cohere released North Mini Code, a 30B-parameter sparse MoE coding model with 3B active parameters, Apache 2.0 weights, and training focused on agentic software engineering workflows.
Cohere North Mini Code is an open coding model aimed at agentic software engineering. Cohere describes it as a 30B-parameter sparse Mixture-of-Experts model with 3B active parameters, Apache 2.0 licensing, 256K total context, and 64K maximum generation. The release matters for builders who want a coding-agent model they can test locally, deploy privately, or run through agent harnesses without depending only on closed hosted models.
Key takeaways
- North Mini Code is Cohere's first open-source model built specifically for developer and coding-agent workflows.
- Cohere says the model has 30B total parameters, 3B active parameters, 256K total context, and 64K maximum generation.
- The weights are available on Hugging Face under Apache 2.0, with BF16 and FP8 variants referenced in the technical post.
- Cohere positions the model for code generation, terminal tasks, repository understanding, code review, and sub-agent orchestration.
- The Hugging Face post gives more detail on training: multi-stage SFT, RLVR, containerized coding environments, and cross-harness robustness.
Practical LinkLoot angle
North Mini Code is worth tracking if your agent stack needs a small-enough coding model with open weights and long-context behavior. The strongest use case is not replacing every frontier coding model; it is testing a controllable baseline for private repositories, terminal-heavy tasks, and repeatable evaluation runs.
| Option | Best use | Limitation | Source |
|---|---|---|---|
| North Mini Code | Open-weight coding-agent experiments and private deployment tests | Cohere-reported benchmarks still need local reproduction | Cohere, Hugging Face |
| Closed hosted coding models | Highest convenience and managed scale | Less control over weights, infrastructure, and offline testing | Vendor docs |
| General open LLMs | Broad local experimentation | Often weaker on terminal and repository-agent behavior | Hugging Face model cards |
For LinkLoot readers, the workflow is clear: benchmark North Mini Code on a known repository task set before using it in production. Include one code-editing task, one terminal task, one code-review task, and one long-context repository question. Track not only pass/fail, but also tool-call validity, repeated loops, diff size, and how often a human has to rescue the run.
What to verify before you act
Verify the exact model artifact, license, and quantization you plan to use on Hugging Face. Reproduce at least a small benchmark in your own harness because Cohere's post combines public scores, internal tests, and harness-specific details. Also check infrastructure fit: a sparse 30B model with 3B active parameters can be efficient, but memory, serving stack, context length, and tool-call formatting still determine whether it works in your environment.
Source check
Cohere's announcement confirms the release, parameter shape, license, context figures, target use cases, and official availability channels. The Hugging Face technical post corroborates the release and adds training and evaluation details, including SFT, RLVR, SWE-Bench, Terminal-Bench, and cross-harness training notes.
It is Cohere's open coding model for agentic software engineering, released with 30B total parameters and 3B active parameters.
For more implementation patterns, see LinkLoot's guide to AI workflow automation.