Cohere released North Mini Code, a 30-billion parameter mixture-of-experts model with 3 billion active parameters designed specifically for agentic software engineering tasks.

The model is available on Hugging Face under the Apache 2.0 license, marking Cohere's first model targeted at developers rather than general language tasks.

North Mini Code achieved a score of 33.4 on Artificial Analysis' Coding Index, outperforming similarly-sized models including Qwen3.5, Gemma 4, and Devstral Small 2. The model also beat substantially larger models such as Nemotron 3 Super (120B parameters) and Mistral Small 4 (119B parameters).

Architecture and Training

The model uses a decoder-only Transformer architecture with 128 experts, activating 8 per token. It employs interleaved sliding-window attention with RoPE and global attention in a 3:1 ratio.

Cohere trained the model using a two-stage supervised fine-tuning process followed by reinforcement learning with verifiable rewards (RLVR). The first stage used 64K context length with code datasets comprising 70% of trainable tokens. The second stage extended to 128K context length using 4.5 billion tokens from high-quality agentic and reasoning samples.

The training pipeline relied on over 70,000 verifiable tasks across approximately 5,000 unique repositories. Cohere deduplicated against SWE-Bench and SWE-Bench-Pro repository sources to prevent evaluation contamination.

"We trained North Mini Code using multiple scaffolds rather than optimizing for a single one," the company said. "This approach enables North Mini Code to serve as a reliable foundation for coding agents."

The model targets complex software engineering workflows, terminal-based agentic tasks, and high-quality code generation. Cohere positioned it as optimized for real-world coding agents that require robustness across different agent frameworks.

Cohere plans to integrate North Mini Code into OpenCode, its coding agent platform, demonstrating the model's practical applications in autonomous software development workflows.