Cohere releases North Mini Code: 30B MoE, 3B active, 256K context

• cohere's launch thread introduced North Mini Code as Cohere's first open-source coding model, with a 30B-parameter MoE architecture, 3B active parameters, and an Apache 2.0 license.
• According to JayAlammar on specs, the model ships with a 256K context window, while Official Cohere release notes position it for agentic coding work rather than plain code completion.
• Artificial Analysis' writeup says North Mini Code scored 33.4 on its Coding Index, which matches the benchmark number Cohere highlighted in the launch thread.
• opencode's day-one support post made the model immediately usable in a coding-agent client the same day the weights landed.

You can jump from Cohere's launch post to the Hugging Face model card, which quietly adds a 64K max output limit and architectural details. Artificial Analysis' benchmark note places the model just above gpt-oss-20B on its Intelligence Index, and opencode's post means this was not just a weights drop, it was usable in an agent client on day one.

Cohere is framing North Mini Code as a small, open coding model for developers, not a frontier-scale general model. The core spec from cohere's thread and JayAlammar on specs is simple:

• 30B total parameters
• 3B active parameters
• 256K context window
• Apache 2.0 license
• Weights released on Hugging Face

The official blog post also calls it Cohere's first agentic coding model and the first member of its next-generation North family.

Cohere's headline number was a 33.4 score on the Artificial Analysis Coding Index, cited in both cohere's launch thread and Artificial Analysis' own writeup. That same Artificial Analysis note also puts the model at 27.6 on its broader Intelligence Index, above gpt-oss-20B at 24.5 and just below Mistral Small 4 at 27.8.

The Hugging Face model card adds the more interesting implementation details:

• decoder-only sparse MoE
• 128 experts, 8 activated per token
• sliding-window and global attention mixed in a 3:1 ratio
• 64K max output, separate from the 256K context limit

Those details did not make the tweet thread, but they matter more than the launch slogan.

The cleanest part of this launch was distribution. opencode's post made North Mini Code free inside OpenCode on the same day, and JayAlammar on specs explicitly called out compatibility.

That gives the release a second story beyond open weights: Cohere shipped a small coding MoE, and an agent client wired it up immediately. nickfrosst's reaction leaned into the local, sovereign angle, calling the model small, cost-effective, Apache 2.0, and locally deployable.