Cohere just dropped its first open-source coding model — and it runs on a single GPU. North Mini Code is a 30B parameter Mixture of Experts model with only 3B active parameters, designed specifically for agentic software engineering. At a time when frontier coding models demand clusters of H100s, North Mini Code brings serious coding capability to developers who want to run models locally or on-premise.
What Is North Mini Code?
North Mini Code is Cohere's inaugural entry in the "North" family of code agent models. It's built around a Mixture of Experts (MoE) architecture — 30B total parameters, but only 3B activate during any given forward pass. This means you get access to the knowledge capacity of a larger model while keeping inference costs and hardware requirements dramatically lower.
The model is released under Apache 2.0, making it fully open-source for commercial and personal use. Weights are available on HuggingFace in three formats: bf16, fp8, and w4a16.
Key Specifications
| Specification | Value |
|---|---|
| Architecture | Mixture of Experts (MoE) |
| Total Parameters | 30B |
| Active Parameters | 3B |
| Context Length | 256K tokens |
| Max Generation | 64K tokens |
| License | Apache 2.0 |
| Minimum Hardware | 1× H100 (FP8 or FP4) |
| Output Speed | ~199 tokens/sec (Cohere API) |
Benchmark Performance
On the Artificial Analysis Coding Index — which weights Terminal-Bench Hard and SciCode scores — North Mini Code achieves 33.4. That places it above models like GLM-4.7-Flash (25.9) and within striking distance of Qwen3.6 35B A3B (35.2), despite having a smaller active parameter count.
On the broader Artificial Analysis Intelligence Index, it scores 27.6, putting it above gpt-oss-20B (24.5) and just barely below Mistral Small 4 at 27.8 — a model with nearly 4× the total parameters and over 2× the active parameters.
Where It Excels
- Software engineering tasks: Strong performance on SWE-Bench-style evaluations
- Terminal tasks: Competitive on Terminal-Bench Hard
- Code generation: Excels at complex, multi-step code generation
- Throughput: 2.8× higher output throughput than Devstral Small 2 under identical conditions
Where It Struggles
North Mini Code is purpose-built for coding. On non-coding agentic tasks, performance drops significantly:
- GDPval-AA: 14% (general-purpose agentic evaluation)
- τ²-Bench Telecom: 37% (telecom domain agent tasks)
- Overall Agentic Index: 21.7
This is a coding specialist, not a generalist. If you need a model for broad agentic workflows, look elsewhere. If you need a model that writes, reviews, and debugs code efficiently — this is targeted at exactly that use case.
Why This Matters
Sovereign AI for Developers
Cohere is positioning North Mini Code as part of their "sovereign AI" push — giving developers and enterprises the ability to run capable coding models on their own infrastructure without vendor lock-in. Apache 2.0 licensing means no API dependencies, no per-token costs, and full control over the model.
The minimum hardware requirement of a single H100 (at FP8 or FP4 quantization) puts this within reach of well-equipped developers and startups — not just large enterprises.
The MoE Efficiency Play
The 3B active parameter count is the real story. Traditional dense models at 30B parameters would require significantly more compute per token. By using MoE, North Mini Code delivers:
- Faster inference — only 3B parameters computed per token
- Lower memory bandwidth — critical for edge and on-prem deployment
- Competitive quality — the full 30B knowledge base is available, just selectively activated
The 256K context window is also notable for an open model. Most open-source alternatives in this size class top out at 32K–128K. This allows working with substantial codebases without chunking or context management tricks.
Getting Started
North Mini Code is available through multiple channels:
- HuggingFace: CohereLabs/North-Mini-Code-1.0 (bf16, fp8, w4a16 variants)
- Cohere API: Free tier available with API key
- Model Vault: Cohere's managed inference environment
- OpenRouter: Available through the routing platform
- OpenCode: Integrated into the OpenCode harness
For local deployment, the w4a16 quantized variant offers the smallest footprint while maintaining reasonable quality. The fp8 version is recommended if you have an H100 available.
How It Compares (June 2026)
| Model | Total Params | Active Params | Coding Index | License |
|---|---|---|---|---|
| North Mini Code | 30B | 3B | 33.4 | Apache 2.0 |
| Qwen3.6 35B A3B | 35B | 3B | 35.2 | Apache 2.0 |
| GLM-4.7-Flash | ~30B | ~3B | 25.9 | MIT |
| Devstral Small 2 | 24B | 24B | ~28 | Apache 2.0 |
| Mistral Small 4 | 119B | 6.5B | N/A | Apache 2.0 |
North Mini Code occupies a compelling middle ground — not the absolute best in its class, but competitive enough to be a practical daily driver for developers who want open-source, on-prem coding assistance without massive infrastructure.
The Bottom Line
Cohere North Mini Code represents a meaningful step forward for open-source coding models. It proves that you don't need frontier-scale infrastructure to get capable agentic coding performance. The Apache 2.0 license, reasonable hardware requirements, and genuine coding capability make it worth evaluating for any team considering local or sovereign AI coding tools.
The model is specialized — it's not going to replace your general-purpose LLM. But as a dedicated coding assistant that you control, it's one of the most efficient options available in the open-source landscape as of June 2026.