Cohere released North Mini Code on 9 June 2026, its first model aimed specifically at developers. The open-weights coding model uses a sparse mixture-of-experts design with 30 billion parameters in total and 3 billion active during each token, targeting code generation, software-engineering agents and terminal work. Its weights are available under Apache 2.0.

Cohere Watch analysis
What happened
The model carries a 256,000-token context window and can generate up to 64,000 tokens. Cohere's model card describes 128 experts, with eight selected per token, plus interleaved local and global attention. That combination is intended to keep inference requirements below those of a dense model with the same total parameter count while retaining enough context for large repositories and long agent runs.
Cohere offers the weights in full-precision and reduced-precision variants. Its launch table lists one H100 accelerator as the minimum hardware for FP8 or FP4 deployment. That is genuinely compact by data-centre standards, although it is not the same thing as running casually on any laptop someone finds beneath the sawdust.
Why it matters
A capable coding model that fits on one accelerator gives engineering teams another way to keep repositories, prompts and tool traces inside their own environment. It also makes a two-tier setup more plausible: use a local specialist for routine edits and reserve larger hosted agents for the jobs where their extra capability earns its keep. The licence permits inspection and adaptation, but operational quality still depends on the serving stack, quantisation and the agent harness around it.
Our read
Cohere reports up to 2.8 times the output throughput of Devstral Small 2 and a 30 percent inter-token-latency advantage in its own tests. Those figures are useful launch evidence, not an independent verdict. The interesting test is less glamorous: whether the model can review, repair and navigate ordinary private repositories all week without its human supervisor reaching for the expensive cloud spanner every afternoon.
What to watch
- Independent repository-level evaluations using repeatable agent harnesses.
- Quality retained by FP8, FP4 and community quantisations.
- Memory use, throughput and stability during long tool-driven sessions.
- Whether teams actually split routine and frontier coding work across local and hosted models.
Discussion spark: Could locally deployed coding models become the default for routine repository work while frontier agents handle only the hardest tasks?
Sources and evidence
- Introducing North Mini Code: Cohere's first model for developers (9 June 2026, 00:00 UTC)
- North Mini Code 1.0 model card (9 June 2026, 00:00 UTC)
not affiliated with or endorsed by Cohere