Sebastian Raschka • 6/12/2026

North Mini Code and Agentic Coding Benchmarks

This article reviews Cohere's North Mini Code, a 30B-parameter Mixture-of-Experts model with 3B active parameters for agentic coding. It highlights the architecture (128 experts, 8 active per token, interleaved attention) and evaluation on Terminal-Bench, SWE-Bench, SciCode, and LiveCodeBench. The model excels in workflow-heavy agentic coding tasks but is competitive but not leading in traditional code generation. Benchmark caveats about harness details and tool APIs are noted.

0 comments

#Agentic Coding #Swe Bench #Mixture Of Experts