Cohere North Mini Code — 3B Active Params Crushes Hard LeetCode

A coding model with 30B parameters that runs on a single GPU and only activates 3B at a time — Cohere just open-sourced it under Apache 2.0. This is North Mini Code, and I have it running live.

Architecture

• Mixture-of-Experts: 30B total, 3B active per token (FP8 quantized) — large-model quality, small-model speed and memory
• 256K context window, 64K max output — headroom for whole codebases
• Tops the sub-5B-active category on SWE-bench Verified and SWE-bench Pro; beats models with twice the active params on Terminal-Bench

Live demo

Four Hard LeetCode problems on a RunPod H100 via vLLM: LRU Cache, Trapping Rain Water, Course Schedule II, Serialize Binary Tree.

Stack: CohereLabs/North-Mini-Code-1.0-fp8 (HF) · vLLM · RunPod H100 Blog → https://cohere.com/blog/north-mini-code

Full breakdown + live runs in the video above. Subscribe for more local-LLM deep dives.