Cohere’s 30B Coding Agent Beats Models 4x Its Size on One H100 — and It Shouldn’t
Last Updated on June 22, 2026 by Editorial Team
Author(s): Chew Loong Nian – AI ENGINEER
Originally published on Towards AI.
Cohere's 30B Coding Agent Beats Models 4x Its Size on One H100 — and It Shouldn't
A 30-billion-parameter model with only 3 billion active parameters has no business landing 0.6 points behind Claude Opus 4.6 on the hardest real-world coding benchmark we have. It has even less business doing it on a single H100 while charging $0 per token. But that is exactly what Cohere’s North Mini Code does, and I have spent the last two days trying to figure out where the catch is.
Summary of the article: After introducing North Mini Code 1.0 and its release details, the article explains the model’s efficiency-first design—especially its sparse Mixture-of-Experts setup (3B active out of 30B total) and its attention architecture—which enables strong performance on a single GPU. It then dissects benchmark results and clarifies the “beats 120B” claim using pass@10 vs. pass@1 (showing big gains on best-of-10 and curated/indexed comparisons, but less dominance for single-shot accuracy on harder benchmarks). The author highlights practical benefits like throughput (tokens per second per dollar), provides step-by-step deployment paths (free OpenRouter endpoint, vLLM self-hosting with a required reasoning parser, or running locally via Ollama), and discusses failure modes: the model excels at local, scoped reasoning but is weaker at long-horizon, multi-file codebase orchestration. The piece concludes that North Mini Code is the most sensible default for an open, efficient coding agent—especially for teams that can’t or won’t send code externally—while acknowledging that frontier or other larger specialists can still win on peak single-shot capability.
Read the full blog for free on Medium.
Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.
Published via Towards AI
Towards AI Academy
We Build Enterprise-Grade AI. We'll Teach You to Master It Too.
15 engineers. 100,000+ students. Towards AI Academy teaches what actually survives production.
Start free — no commitment:
→ 6-Day Agentic AI Engineering Email Guide — one practical lesson per day
→ Agents Architecture Cheatsheet — 3 years of architecture decisions in 6 pages
Our courses:
→ AI Engineering Certification — 90+ lessons from project selection to deployed product. The most comprehensive practical LLM course out there.
→ Agent Engineering Course — Hands on with production agent architectures, memory, routing, and eval frameworks — built from real enterprise engagements.
→ AI for Work — Understand, evaluate, and apply AI for complex work tasks.
Note: Article content contains the views of the contributing authors and not Towards AI.
Related posts
Recent Posts
