North Mini Code: 30B Sparse MoE Coding Model

Model Overview

North Mini Code (1.0) is a specialized 30B total parameter, sparse Mixture-of-Experts (MoE) coding model with ~3B active parameters per token developed by Cohere and Cohere Labs.

Designed for agentic software engineering and developer workflows, it achieves near-30B-scale reasoning and multi-file code editing performance with the low compute overhead of a 3B parameter model.

---

Key Features

• Sparse MoE Architecture: 30B total parameters across 128 experts with 8 active per token (~3B active parameters).
• Hybrid Attention Design: Interleaves sliding-window attention (with RoPE) and global attention in a 3:1 ratio.
• Extended Context Capability: Features a 256K token context window with up to 64K output token generation.
• Agentic Optimization: Post-trained via two-stage SFT and RLVR for multi-file repo changes, terminal execution, and scaffolds like OpenCode and SWE-Agent.
• Open Weights & Deployability: Released under the permissive Apache 2.0 license for single-GPU local execution.

---

Verified Project Links

• Developer Blog: https://cohere.com/blog/north-mini-code
• GitHub Repository: https://github.com/cohere-ai
• Hugging Face Model: https://huggingface.co/CohereLabs/North-Mini-Code-1.0

---

Performance & Benchmarks

• SWE-bench Verified: 67.6% (pass@1) / 80.2% (pass@10).
• HumanEval: 50.0%.
• Artificial Analysis Coding Index: 33.4.