North Mini Code and Agentic Coding Benchmarks

This article covers Cohere's release of North Mini Code, a 30B-parameter Mixture-of-Experts model (3B active parameters, 128 experts, 8 active per token) under Apache 2.0, designed for agentic coding tasks. It details the model's architecture, including interleaved sliding-window and global attention, and its evaluation on agentic benchmarks like Terminal-Bench (terminal interaction), SWE-Bench (GitHub issue resolution), and traditional coding benchmarks (SciCode, LiveCodeBench). The article highlights the model's strong performance on workflow-heavy tasks compared to Gemma 4 and Qwen3.6, while noting that benchmark results depend heavily on evaluation setup details. It is a tech-focused analysis of a new AI coding model.