North Mini Code: Cohere's First Model for Developers (huggingface.co)

Cohere has introduced North Mini Code, a pioneering 30B-parameter Mixture-of-Experts model optimized for agentic coding tasks, now available on Hugging Face under the Apache 2.0 license. This model stands out with its ability to engage in complex software engineering workflows by deploying 3B active parameters, achieving impressive scores on the Artificial Analysis Coding Index. Notably, it surpasses both smaller and larger models, demonstrating its robustness and effectiveness in generating high-quality code. North Mini Code's architecture features an innovative mixture of sliding-window and global attention mechanisms and employs a unique two-phase post-training refinement with supervised fine-tuning followed by reinforcement learning with verifiable rewards, focusing on both traditional coding and terminal-based tasks. The significance of North Mini Code lies in its capacity to enhance productivity in software development by providing a reliable foundation for coding agents like OpenCode. Its versatility is further highlighted by its robust performance across various coding harnesses, enabling cross-harness generalization while maintaining benchmark performance. By utilizing a multi-environment reinforcement learning approach, the model not only improves accuracy but also reduces common issues such as errors and hallucinations during code generation. This development marks a substantial advancement in the usability of AI models for real-world coding applications, potentially transforming the landscape of automated software engineering.