North Mini Code 1.0

North Mini Code 1.0, published by Cohere and Cohere Labs on 5 June 2026, is a sparse mixture-of-experts model built for one job: writing code, working through software-engineering tasks and driving a terminal. It holds 30.48 billion parameters in total but activates roughly 3 billion per token, drawing 8 of its 128 experts each time — the arrangement that lets a model of this size run on hardware sized for a much smaller one. The context window is unusually wide for a model of this class: 256,000 tokens in, with up to 64,000 tokens of output, which is enough to hold a mid-sized repository and still have room to write a patch. Attention is mixed rather than uniform — of 49 layers, most use a 4,096-token sliding window and only every fourth uses full attention, which is what keeps the long window affordable. Two behaviours matter in practice. The model supports interleaved thinking and, as Cohere puts it, works best with it switched on: the reasoning it produces between tool calls is meant to be passed forward into later steps rather than discarded, which is a real constraint on how an agent harness has to be written around it. And tool use is native, driven by JSON-schema function definitions through the chat template. The licence is plain Apache 2.0 with no thresholds, no revenue clauses and no separate agreement for commercial use — the same terms Cohere gave its Command A+ weights, and a contrast to the bespoke licences increasingly attached to flagship open releases. Cohere benchmarks the model on SWE-bench Verified, SWE-bench Pro, Terminal-Bench v2 and Hard, SciCode and LiveCodeBench v6, averaging three seeds per run, but publishes the results only as a chart image rather than a table, so we do not quote figures we cannot read from the source.

▸Videos

No videos yet.

▸Reviews