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

Cohere introduces North Mini Code, a 30B-parameter Mixture-of-Experts model designed for agentic software engineering tasks.

• •North Mini Code achieves a Coding Index score of 33.4, outperforming larger models like Nemotron 3 Super (120B) and Mistral Small 4 (119B) in coding benchmarks
• •The model uses two-stage supervised fine-tuning followed by reinforcement learning with verifiable rewards (RLVR), with code data comprising 70% of training tokens
• •Training leveraged over 70,000 verifiable tasks across 5,000 unique repositories with sample-level filtering to ensure quality
• •Cross-harness robustness was achieved by exposing the model to diverse agent harnesses during second-stage SFT with only 6% benchmark harness data
• •Asynchronous RL training decouples vLLM sampler and learner processes using CISPO objective to efficiently handle variable-length coding agent rollouts