Cohere releases North Mini Code, an open-weight agentic coding model

Category: AI
Status: developing
Sources: Cohere blog
Summary: Cohere published North Mini Code 1.0 on 2026-06-09, a mixture-of-experts coding model with 30B total and 3B active parameters, a 256K context window, and up to 64K generation. It is released under Apache 2.0 and is available on Hugging Face, the Cohere API, Model Vault, and OpenRouter. Cohere reports a 33.4 score on the Artificial Analysis Coding Index and, in internal testing, up to 2.8x higher output throughput and a 30 percent inter-token-latency advantage over Devstral Small 2, with evaluation on SWE-Bench Verified, SWE-Bench Pro, Terminal Bench v2, and Terminal Bench Hard using SWE-agent and ReAct harnesses. Minimum stated hardware is a single H100 at FP8. The throughput and benchmark figures are the vendor's own and are not independently reproduced.
Why it matters: A permissively licensed 30B/3B MoE coding model that runs on one H100 adds to the open-weight pressure (alongside GLM 5.2 and Kimi K2.7-Code) on teams paying premium per-token rates for proprietary coding agents.
Follow-up: Watch for independent coding-benchmark results and real-world agent-harness adoption.