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

Cohere Releases North Mini Code for Agentic Development

Cohere has released North Mini Code, its first model from a new family designed specifically for developers. The 30B-parameter Mixture-of-Experts (MoE) model, with 3B active parameters, is now available on Hugging Face under the permissive Apache 2.0 license. Unlike general-purpose coding assistants, North Mini Code is optimized for complex, agentic software engineering workflows, positioning it as a foundational tool for building autonomous code agents. Its release marks a focused effort to address the practical challenges of integrating AI into real-world software development cycles.

Technical Design and Training Pipeline

The model's architecture and training process underscore its focus on performance and efficiency for agentic tasks. North Mini Code is built on a decoder-only Transformer foundation that leverages a sparse MoE structure to manage its large parameter count. Its post-training regimen is particularly notable, employing a multi-stage process to instill robust coding and reasoning capabilities.

• Architecture: A sparse Mixture-of-Experts (MoE) model with 128 experts, activating 8 per token.
• Attention: Uses an efficient implementation that interleaves sliding-window self-attention with global self-attention.
• Phase 1 - SFT: Initial supervised fine-tuning on a broad mix of programming, reasoning, and instruction-following data.
• Phase 2 - SFT: A second, more targeted SFT stage using a high-quality data mixture from verified agentic and reasoning-driven samples.
• Phase 3 - RLVR: A final phase of agentic reinforcement learning with verifiable rewards (RLVR) to refine performance on software engineering and terminal tasks.

A Focus on Multi-Harness Robustness

A key differentiator for North Mini Code is its explicit training for robustness across diverse agent harnesses. Recognizing that real-world AI agents must operate in varied tooling environments, Cohere trained the model on data from multiple scaffolds, including SWE-Agent, mini-SWE-agent, and OpenCode. This approach improves the model’s ability to generalize its skills, rather than being over-optimized for a single benchmark or toolset. The company also implemented an asynchronous reinforcement learning loop, which decouples data sampling from the training process. This technical choice addresses the high variability in coding task completion times, improving training throughput and stability for long, complex agentic rollouts.

Strategic Takeaway: Cohere's strategy with North Mini Code signals a market shift from pure benchmark performance to the operational robustness of AI agents, focusing on multi-harness generalization and efficient training methodologies to address real-world software engineering complexities.