Introducing North Mini Code: Cohere’s First Model For Developers

2026-06-11 · Source: Hugging Face - Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Advanced, long

Summary

Cohere has released North Mini Code, a 30B-parameter Mixture-of-Experts model with 3B active parameters, available on Hugging Face under the Apache 2.0 license as of June 9, 2026. This model is specifically engineered for agentic software engineering tasks and complex code generation. North Mini Code achieved a score of 33.4 on Artificial Analysis' Coding Index, surpassing models like Qwen3.5 (35B-A3B) and even larger ones such as Nemotron 3 Super (120B-A12B). Its architecture is a decoder-only Transformer-based sparse MoE, featuring an efficient attention implementation and 128 experts. The model undergoes a two-stage supervised fine-tuning followed by reinforcement learning with verifiable rewards (RLVR), utilizing 64K and 128K context lengths. This training approach enhances robustness across diverse agent harnesses, yielding a 10% gain on OpenCode evaluation and improving Terminal-Bench v2 performance by 7.9% pass@1.

Key takeaway

For AI Engineers evaluating coding models for agentic software engineering, North Mini Code offers a compelling open-source option. Its 30B MoE architecture and specialized RLVR training deliver strong performance, outperforming larger models on key benchmarks. You should consider integrating this model, available on Hugging Face, into your agentic workflows, particularly for complex terminal-based tasks or environments requiring cross-harness robustness.

Key insights

North Mini Code is a 30B MoE model excelling in agentic coding through cascaded SFT and RLVR.

Principles

• Multi-stage training improves robustness.
• RLVR enhances agentic performance.
• Cross-harness training yields positive transfer.

Method

A two-stage SFT (64K then 128K context) primes for RLVR, followed by asynchronous multi-environment online RL using a windowed FIFO queue and CISPO objective.

In practice

• Integrate North Mini Code for agentic software tasks.
• Utilize BF16 or FP8 quantized weights.
• Explore OpenCode for agentic development.

Topics

• North Mini Code
• Mixture-of-Experts
• Agentic AI
• Software Engineering
• Reinforcement Learning
• Code Generation

Code references

• anomalyco/opencode
• SWE-agent/mini-swe-agent
• laude-institute/harbor
• laude-institute/terminal-bench

Best for: AI Architect, Research Scientist, AI Engineer, Machine Learning Engineer, AI Scientist

Related on AIssential

• North Mini Code: Agentic Coding Model for Developers - Cohere — Cohere's North Mini Code is an efficient, open-source agentic coding model designed for sovereign developer control.
• North Mini Code - Cohere — A compact 3B active parameter MoE model excels in agentic coding and local deployment.
• Cohere open-sources a coding agent that runs on a single H100 — Cohere's North Mini Code offers an open-source, locally deployable MoE coding agent for complex software engineering tasks.
• [AINews] Founders and Forward Deployed Engineers — The AI landscape is rapidly converging on integrated agentic systems, driven by both proprietary and increasingly capable open models.
• Introducing Google Antigravity 2.0 — Agentic coding is a necessary but insufficient step towards general model intelligence, requiring co-optimized layers and user-friendly interfaces.
• The Free Open Model Matching GPT-5.5 on Agentic Coding — Nex-N2-Pro, a free open-weights model, matches GPT-5.5's agentic coding performance on SWE-Bench Pro.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Hugging Face - Blog.