North Mini Code Review: Cohere's First Open-Source Coding Model
Cohere launched North Mini Code on June 9, 2026 — a 30B MoE agentic coding model with only 3B active parameters, Apache 2.0 license, and 256K context. Here's what's new, what the benchmarks show, and whether it belongs in your stack.
Reviewed 2026-06-11 · Source: Cohere announcement
Quick Verdict
North Mini Code is a strong entry into the small open-source coding model tier. The MoE architecture (30B total, 3B active) keeps inference costs minimal while delivering competitive benchmark scores — including 2.8× higher throughput than Devstral Small 2 on identical hardware. The Apache 2.0 license, single H100 requirement, and 256K context make this a practical choice for developers who want an open coding model they can self-host without serious infrastructure overhead. For Cohere, it's a meaningful strategic move: they're best known for enterprise search, and North Mini Code plants a flag in the developer and open-source ecosystem.
What's New in North Mini Code
Mixture of Experts at 30B total / 3B active
North Mini Code uses a Mixture of Experts architecture with 30B total parameters but only 3B active per token. This gives the model the capacity of a larger model while keeping inference costs low — comparable to running a 3B dense model for most workloads. The single H100 minimum requirement (at FP8) makes it accessible to individual developers and small teams without multi-GPU infrastructure.
Agentic coding and terminal task focus
North Mini Code is purpose-built for software engineering workflows: code generation, agentic task execution, and terminal operations. Cohere benchmarked it specifically on real-world agentic tasks, not just static code completion. It achieves a 33.4 on the Artificial Analysis Coding Index and outperforms Devstral Small 2 on throughput by 2.8× under identical hardware conditions.
Apache 2.0 open-source license
The model is released under the Apache 2.0 license — the same standard permissive open-source license used by Llama and other fully open models. You can use it commercially, modify it, fine-tune it, and redistribute it without restrictions. Weights are available on Hugging Face (CohereLabs/North-Mini-Code-1.0). This is notably more open than some competitors that use modified or custom licenses.
256K context with 64K max generation
North Mini Code supports 256K tokens of total context — enough for large codebases, long agentic sessions, or bulk file analysis. The 64K max generation length is practical for generating extensive code artifacts in a single pass. For reference, most small coding models cap generation at 4K–8K tokens.
Inaugural model in Cohere's next-gen North family
North Mini Code is the first model in Cohere's new North generation — the company's next-generation model family. Cohere has historically focused on enterprise search and RAG (Command/Embed families), so North Mini Code marks a strategic expansion into developer tooling and open-source model weights. More North models are implied by the 'inaugural member' framing.
Benchmarks
Key performance numbers from Cohere's announcement (June 9, 2026).
| Benchmark | North Mini Code | Notes |
|---|---|---|
| Artificial Analysis Coding Index | 33.4 | Competitive among similarly sized open-source coding models |
| Output throughput vs Devstral Small 2 | 2.8× faster | Identical concurrency levels and hardware configurations |
| Inter-token latency vs Devstral Small 2 | 30% better | Consistency and pacing of token generation |
| Total parameters | 30B | MoE architecture — only 3B active per token |
| Context window | 256K tokens | 64K max generation length |
| Minimum hardware | 1× H100 @ FP8 | Runs on a single GPU — unusually accessible for a 30B MoE |
Pricing & Access
Apache 2.0 license. Download weights from Hugging Face (CohereLabs/North-Mini-Code-1.0). Minimum: 1× H100 @ FP8. Runs in any framework supporting standard LLM weight formats.
Available via the Cohere API with a standard API key. See cohere.com for current per-token pricing. Suitable for cloud-based inference without managing your own GPU.
Deploy in a managed, dedicated inference environment on Cohere Model Vault. Provides guaranteed throughput, private endpoints, and enterprise SLAs.
Available on OpenRouter for routing across providers. Also directly accessible via OpenCode for AI-assisted coding workflows — no Cohere account required.
Who should use North Mini Code?
- ✓Solo developers and small teams — the single H100 requirement means you can self-host a frontier-quality coding model without a multi-GPU cluster. Apache 2.0 means no licensing overhead.
- ✓High-throughput coding pipelines — 2.8× the output throughput of Devstral Small 2 under identical hardware means you can process significantly more coding tasks per hour at the same cost.
- ✓Agentic software engineering — designed for multi-step coding agents, terminal tasks, and extended code generation sessions. 256K context handles large codebases without chunking.
- ✓Organizations requiring data sovereignty — Apache 2.0 weights on a single H100 mean full on-premise deployment with no data leaving your infrastructure.
- ✗General-purpose chat and reasoning — North Mini Code is optimized for code. For broad reasoning, instruction-following, or multimodal tasks, models like Mistral Small 4 or Claude Haiku are better fits.
- ✗Maximum coding benchmark scores — for the absolute best on SWE-Bench and Polyglot, Mistral Medium 3.5 (77.6% SWE-Bench) and Claude Opus 4.8 still lead. North Mini Code competes at the small-model tier.
Frequently Asked Questions
What is North Mini Code?
North Mini Code is Cohere's first agentic coding model, released June 9, 2026 under the Apache 2.0 open-source license. It's a Mixture of Experts model with 30B total parameters and 3B active per token, designed for code generation, agentic software engineering, and terminal tasks. It requires just one H100 GPU at FP8 to run.
How does North Mini Code compare to Devstral Small 2?
North Mini Code outperforms Devstral Small 2 on output throughput by 2.8× and has 30% better inter-token latency under identical hardware configurations. Time-to-first-token is more closely matched, with Devstral Small 2 holding a slight edge there. Both are competitive on coding benchmarks at the small model tier.
Is North Mini Code truly open source?
Yes. North Mini Code is released under the Apache 2.0 license — a standard, permissive open-source license. You can use it commercially, modify it, fine-tune it, and redistribute it without restrictions. The weights are freely available on Hugging Face.
What hardware does North Mini Code require?
The minimum hardware listed by Cohere is one NVIDIA H100 GPU running at FP8 precision. This is notably accessible for a 30B parameter model — comparable small open-source models often require 2–4× A100s or H100s. The MoE architecture (3B active per token) is what makes single-GPU inference practical.
What is the North model family?
North is Cohere's next generation of models, with North Mini Code as the inaugural member. The name suggests more North models are planned. Cohere's prior model families were Command (instruction-following/chat), Embed (embeddings for search), and Rerank. North appears to be their new frontier and developer-focused line.
Where can I try North Mini Code?
You can access North Mini Code via Hugging Face (download weights), the Cohere API (API key required), Cohere Model Vault (dedicated managed inference), OpenRouter (multi-provider routing), or OpenCode (AI coding assistant). The Hugging Face weights are free under Apache 2.0.
Try North Mini Code
Available as open weights on Hugging Face (Apache 2.0), via the Cohere API, and on OpenRouter.
Try North Mini Code →Open weights on Hugging Face · Apache 2.0 · Minimum: 1× H100 @ FP8
Affiliate disclosure: Some links on this page are affiliate links. If you sign up through them, AISO Tools may earn a commission at no extra cost to you. This never affects our rankings or reviews.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.
Join thousands of professionals who discover the best AI tools every week. No spam — unsubscribe anytime.