Cohere has released North Mini Code, a 30-billion-parameter Mixture-of-Experts (MoE) model featuring 3 billion active parameters specifically engineered for agentic software tasks. Available on Hugging Face under an open-source Apache 2.0 license, the model serves as the initial entry in Cohere's new family of developer-focused models. Software engineering teams can deploy this lightweight system across diverse agent harnesses to execute terminal commands and resolve repository-level code issues.
Architecture and MoE Engineering
North Mini Code utilizes a decoder-only Transformer architecture with 128 total experts, activating 8 experts per token through a sigmoid-routed feed-forward network. To streamline long-context processing, Cohere implemented interleaved self-attention, alternating sliding-window attention with Rotary Position Embeddings (RoPE) against global attention in a 3:1 ratio.
The development team trained the system using a two-stage cascaded supervised fine-tuning (SFT) pipeline, expanding context windows from 64K tokens in stage one to 128K tokens in stage two. Benchmark evaluations show that North Mini Code scores 33.4 on Artificial Analysis' Coding Index, outperforming competing systems like Gemma 4 (26B-A4B) and substantially larger models including Nemotron 3 Super (120B-A12B).
Training Pipeline and Harness Versatility
To prevent the model from over-fitting to a single agent interface, Cohere exposed North Mini Code to multiple agent scaffolds during SFT, including SWE-Agent, mini-SWE-agent, and OpenCode. Post-training was capped off with asynchronous Reinforcement Learning with Verifiable Rewards (RLVR) using the CISPO log-likelihood objective alongside vLLM rollout sidecars.
Key performance metrics from the official release evaluation include:
- SWE-Bench Verified: Recorded an 80.2% pass@10 during SFT, with RLVR adding a 3.0% absolute gain to pass@1 performance.
- Terminal-Bench v2: Reached a 55.1% pass@10 post-SFT, with RLVR delivering a 7.9% absolute increase in pass@1 scores.
- mini-SWE-Agent: Achieved 61.0% pass@1 through cross-harness generalization.
- Verifiable Dataset: Trained on more than 70,000 executable tasks across approximately 5,000 unique software repositories.
Operational Considerations
For engineering teams deploying North Mini Code within autonomous execution environments, step calibration requires careful oversight. Cohere identified a explicit trade-off during RLVR training: granting the agent a turn budget significantly larger than required leads to degraded trajectory efficiency, prompting unnecessary verbosity and erratic action cycles during execution.