Cohere Launches North Mini Code: An Open Source 30B Agentic Coding Model for Developers

AIDeveloper44 Team

Cohere Enters the Open-Source Developer Ecosystem

On June 9, 2026, Cohere officially announced the release of North Mini Code, the company's first open-source agentic coding model. Released under the highly permissive Apache 2.0 license, North Mini Code is designed specifically to power the next generation of sovereign developer tools. The launch signifies a strategic shift for Cohere, a company traditionally known for enterprise-focused proprietary models like Command A+. By embracing the open-source community, Cohere is directly answering the growing demand for local, private, and highly capable software engineering agents that operate without vendor lock-in.

Mixture-of-Experts Architecture for Maximum Efficiency

Under the hood, North Mini Code utilizes an advanced sparse Mixture-of-Experts (MoE) architecture. The model boasts 30 billion total parameters, but efficiently activates only 3 billion parameters during inference. This sparse activation model strikes a careful balance between robust reasoning capabilities and computational efficiency.

Despite its vast knowledge base and strong capabilities, the model is engineered to require minimal hardware. Specifically, North Mini Code can operate efficiently on a single H100 GPU at FP8 precision. It has also been demonstrated to run smoothly on local setups, such as consumer-grade Apple Mac Studios, consuming around 20 gigabytes of unified memory. Key specifications include:

• Massive Context Window: A 256K total context length paired with a 64K maximum generation limit, allowing it to ingest and navigate entire complex code repositories in a single pass.
• MoE Design: 128 distinct routing experts with 8 activated per token, keeping the compute footprint incredibly lean while maintaining high accuracy.
• Rigorous Training Methodology: The model underwent extensive two-stage supervised fine-tuning across over 70,000 verifiable tasks spanning roughly 5,000 repositories, followed by Reinforcement Learning with Verifiable Rewards (RLVR).

Agentic Coding Capabilities and Benchmarks

Unlike standard text or code-completion LLMs, North Mini Code was built from the ground up specifically for agentic software engineering. It excels at multi-step tasks such as orchestrating sub-agents, mapping complex systems architecture, understanding inter-file dependencies, and executing comprehensive code reviews.

According to public benchmarks and Artificial Analysis, North Mini Code achieved a highly competitive score of 33.4 on the Coding Index. It consistently outperforms similarly sized dense models and holds its own against much larger counterparts. It scored an impressive 80.2% pass@10 and 61.0% pass@1 on SWE-Bench Verified using the mini-SWE-Agent harness, surpassing models like Devstral 2 (123B) and Nemotron 3 Super.

Speed and Latency Advantages

In real-world agentic workflows, speed is just as critical as accuracy. Cohere's internal testing revealed that North Mini Code delivers exceptional throughput for developer tasks. The model outperformed Mistral's Devstral Small 2 (a 24B dense model) by delivering up to 2.8x higher output throughput under identical concurrency levels and hardware configurations.

Additionally, North Mini Code showcased a 30% advantage in inter-token latency. This metric directly impacts the consistency and perceived speed of text generation, making it feel exceptionally snappy for interactive terminal tasks and autonomous debugging workflows. While Devstral Small 2 maintained a slight edge in Time-to-First-Token (TTFT), North Mini Code's sustained generation speed makes it highly optimal for large refactoring jobs that require producing thousands of lines of output.

Sovereign AI: Freedom from Vendor Lock-In

Cohere's decision to open-source North Mini Code highlights a growing industry movement toward sovereign AI. Engineering teams building agentic coding pipelines now have a concrete open-source alternative to managed frontier models like Claude Fable 5 or GPT-5.1 Codex. By releasing the model weights on Hugging Face and ensuring out-of-the-box compatibility with open developer harnesses like OpenCode, SWE-agent, and Terminus-2, Cohere empowers developers to maintain complete oversight over their agentic infrastructure.

For those who prefer managed infrastructure, the model is also available via Cohere's fully managed inference platform, Model Vault, as well as the standard Cohere API and OpenRouter platforms.

The Bottom Line

North Mini Code represents a major milestone for software developers who require enterprise-grade, agentic coding capabilities without the financial, privacy, or network constraints of proprietary cloud APIs. By delivering best-in-class throughput, deep repository understanding, and practical local deployability, Cohere's newest model is set to accelerate the adoption of autonomous coding agents across the entire software industry. Developers can now build, iterate, and deploy AI assistants strictly on their own terms.

Enjoyed this?

Get more posts like this delivered to your inbox.

Subscribe