Cohere's North Mini Code: Open-Source 30B MoE Model Dominates Agentic Coding

Cohere has released North Mini Code, its first model tailored for developers and agentic coding tasks. This 30-billion-parameter Mixture-of-Experts (MoE) transformer uses only 3 billion active parameters per inference, yet achieves a score of 33.4 on the Artificial Analysis Coding Index—outperforming much larger models like Nemotron 3 Super (120B-A12B) and Mistral Small 4 (119B-A6B). The model is available on Hugging Face under the Apache 2.0 license, making it a strong open-source contender for building coding agents. North Mini Code is designed for complex software engineering workflows, terminal-based agentic tasks, and high-quality code generation, offering developers a cost-efficient alternative to dense models of similar capability.

Architecturally, North Mini Code employs 128 experts with 8 activated per token, interleaving sliding-window and global self-attention in a 3:1 ratio. Its post-training pipeline is particularly innovative: a two-stage supervised fine-tuning (SFT) followed by reinforcement learning with verifiable rewards (RLVR) focused on agentic coding. The first stage uses 70% code tokens (43% agentic tool-use, 27% single-turn coding), while the second stage refines with 4.5B tokens of high-quality agentic data. The model was trained on over 70,000 verifiable tasks from roughly 5,000 unique repositories, with careful deduplication against SWE-Bench sources to avoid evaluation leakage. Context lengths start at 64K then scale to 128K via a 'long-to-longer' cascade, ensuring robust performance on both short and extended coding tasks. For developers building AI-assisted programming tools or autonomous code agents, North Mini Code provides a state-of-the-art foundation that balances efficiency with raw performance.