Cohere's first developer-focused model and the first in its new North family — an open-weights (Apache 2.0) 30B-A3B MoE built for agentic software engineering and terminal tasks. 256K context, 64K max output. Attention interleaves sliding-window (RoPE) with global no-positional-embedding layers in a 3:1 ratio; the MoE block routes top-8 of 128 SwiGLU experts through a sigmoid router, with a single dense layer before the sparse stack.

Post-trained with two-stage cascaded SFT followed by agentic RLVR targeting software engineering and terminal tasks, deliberately trained across multiple agent scaffolds (positioned as a foundation for coding agents like OpenCode) rather than tuned to one harness. Cohere reports 33.4 on AA's Coding Index, ahead of similar-size open models (Qwen3.5 35B-A3B, Gemma 4 26B-A4B, Devstral Small 2) and some far larger ones (Nemotron 3 Super 120B-A12B, Devstral 2 123B). AA Intelligence Index v4.1: 21. Benchmarked on SWE-Bench Verified/Pro, Terminal-Bench v2/Hard, SciCode, and LiveCodeBench v6. "North" extends the name of Cohere's enterprise agent platform to a model family.

Model Details

Architecture MOE
Parameters 30B
Active params 3B
Experts 128 (top-8)
Context window 262,144
AA Intelligence 21
License Apache 2.0
codingagenticopen-weightmoe