NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·1h ago
Data Science
Sebastian Raschka Blog·Jun 12·5 min read
North Mini Code and Agentic Coding Benchmarks
TL;DR
Software engineers face challenges in coding tasks that require interaction with tools rather than just generating code. Cohere developed North Mini Code, a 30B-parameter Mixture-of-Experts model designed for agentic coding tasks.
✦ Why It Matters
Engineers can utilize North Mini Code to improve coding efficiency in tool-integrated environments.
Key Takeaways
- Explore North Mini Code for agentic coding applications.
- Consider the impact of tool APIs on model performance.
- Evaluate models based on workflow-heavy task benchmarks.
- Stay updated on evolving coding benchmarks and methodologies.
Full Summary
Agentic coding tasks require models to interact within a tool loop, which is a gap in traditional code generation approaches. Cohere's North Mini Code is a 30 billion parameter Mixture-of-Experts model that activates 3 billion parameters, featuring 128 experts with 8 active experts per token.
It employs a unique architecture combining interleaved sliding-window and global attention mechanisms. The evaluation setup emphasizes the model's performance in agentic coding, where it significantly outperforms Gemma 4 in workflow-heavy tasks.
While it remains competitive in traditional code generation, it does not reach the performance level of Qwen3.6. These results are based on benchmarks as of June 12, 2026, and highlight the importance of harness details, tool APIs, and prompt templates in influencing outcomes.
Engineers and researchers can leverage these insights to enhance their coding workflows and model evaluations.