Cohere Unveils North Mini Code, a Coding Model for Agents

Cohere has released North Mini Code, designed for developer and agent workflows, under the Apache 2.0 license. This model adopts a 30B-parameter Mixture-of-Experts (MoE) architecture, and maximizes efficiency by using only 3B active parameters during inference.

Key Features and Performance:

• Agent Optimization: Designed specifically for terminal-based tasks and complex software engineering workflows.
• Overwhelming Benchmark Performance: It scored 33.4 points on Artificial Analysis's Coding Index, outperforming comparable open-source models such as Qwen3.5(35B-A3B) and Gemma 4(26B-A4B).
• Architecture: Uses an efficient Transformer decoder structure that interleaves sliding-window attention and global attention at a 3:1 ratio.
• Advanced Training Pipeline: Enhanced code executability and accuracy through a 2-stage SFT (Supervised Fine-Tuning) process and RLVR (Reinforcement Learning with Verifiable Rewards) using verifiable rewards.