mixture-of-experts agentic-coding code-agents open-source

Cohere releases North Mini Code for agentic tasks

30B MoE model with 3B active parameters trained on 70k verifiable coding tasks across containerized environments, optimized for cross-harness agent reliability rather than single-benchmark performance.

June 10, 2026

Summary

Agents built on single-harness-optimized models break when tooling changes (CLI vs JSON vs text). North Mini Code's multi-harness post-training reduces the friction of deploying coding agents across different frameworks without retraining.

Why it matters

Implementation verdict

Replaces smaller coding models (Devstral Small, Gemma 4) for agent workloads where robustness matters more than raw benchmark score. Requires 3B active params (manageable inference cost) and containerized task environments for your own RLVR stage if you need domain specificity. Worth trying now for SWE-Bench and terminal-based tasks; Apache 2.0 licensed on HuggingFace.

Sources

1.30B-parameter Mixture-of-Experts model with 3B active parameters
2.specifically designed and trained for agentic software engineering tasks
3.128 experts, of which 8 are activated per token
4.over 70k verifiable tasks across ~5k unique repositories
5.code datasets correspond to 70% of trainable tokens, 43% agentic tool-use data, and 27% single-turn competitive or scientific programming data
6.achieves 61.0% pass@1 using mini-SWE-Agent
7.10% gain on the evaluation with OpenCode harness while maintaining performance with SWE-Agent

Dev Signal

Get briefs like this in your inbox — free, 3x a week.

100+ sources compressed into one 4-minute read. Ranked, cited, implementation-ready.

Read the full issue →All briefs