North Mini Code Review: Cohere's 30B Open-Source Coding Model — Fast, Free, and Worth Running Locally?

North Mini Code is Cohere's first open-source agentic coding model, released June 9, 2026 under an Apache 2.0 license. It is a 30-billion-parameter mixture-of-experts (MoE) model with just 3B active parameters, designed to run on a single H100 GPU at FP8 while delivering competitive coding performance. The key thing to know: it is not the smartest small model on paper — it scores 27.6 on the Artificial Analysis Intelligence Index, slightly below Qwen3.6 on coding — but it is one of the fastest, hitting up to 2.8x the output throughput of Devstral Small 2 on identical hardware.

Here is the short version for developers deciding whether to deploy it:

• It is genuinely small and cheap to run. 30B total / 3B active means it fits on 1× H100 at FP8 and serves fast — around 199 output tokens per second on Cohere's API, with a 256K-token context window.
• It is fully open and sovereign. Apache 2.0 weights on Hugging Face mean you can deploy on-prem, air-gapped, or locally with zero vendor lock-in or per-token billing.
• It is a coding specialist, not a generalist. It scores 33.4 on the Artificial Analysis Coding Index but only 14% on GDPval-AA and 37% on τ²-Bench Telecom — strong at code, weak at broader agentic reasoning.

This review covers what North Mini Code is, what the benchmarks actually say (including a discrepancy worth flagging), how it compares to Devstral and Qwen, and a direct verdict on when to use it. Source links are embedded throughout so you can verify every number.

What is North Mini Code?

North Mini Code is a 30B-parameter mixture-of-experts model optimized for code generation, agentic software engineering, and terminal tasks. It is the inaugural model in Cohere's new generation of open-weight developer models and the company's first model aimed squarely at developers rather than enterprise workplace tools. You can read the full announcement on Cohere's official blog.

The launch specifications, from the official model card:

• Model: North-Mini-Code-1.0
• License: Apache 2.0 (commercial use permitted, no restrictions)
• Size: 30B total parameters, 3B active per forward pass
• Context length: 256K total context, 64K maximum generation
• Optimized for: code generation, agentic software engineering, terminal tasks
• Minimum hardware: 1× H100 GPU at FP8
• Availability: Hugging Face (BF16 and FP8 weights), Cohere API, Cohere Model Vault, and OpenRouter

The MoE design is the whole story here. Because only 3B of the 30B parameters activate per token, North Mini Code runs at the speed and memory cost of a roughly 3B model while drawing on the knowledge of a 30B one. That is what lets it fit on a single H100 and serve at high throughput — the architectural bet that defines the entire "small but capable" model class in 2026.

Cohere specifically trained it for compatibility with OpenCode, the open-source terminal coding agent, though it works with most coding harnesses.

The benchmarks: what North Mini Code actually scores

North Mini Code's headline numbers are competitive for its size class, but they need context — and one of them deserves scrutiny.

On the Artificial Analysis Coding Index, North Mini Code scores 33.4, which Cohere describes as competitive against similarly sized open models. That score sits significantly above GLM-4.7-Flash (25.9) but below Qwen3.6 35B-A3B (35.2). On the broader Artificial Analysis Intelligence Index v4.0, it scores 27.6 — above gpt-oss-20B at 24.5, and just below Mistral Small 4 (a much larger 119B model) at 27.8.

The picture splits cleanly by task type. North Mini Code is a coding specialist: it holds its own on Terminal-Bench Hard and SciCode (the components of the Coding Index), but it underperforms badly on non-coding agentic work, scoring just 14% on GDPval-AA and 37% on τ²-Bench Telecom, for an overall Agentic Index of 21.7. In plain terms: it writes and edits code well, but it is not the model you want orchestrating a complex multi-domain business workflow.

My own read — a benchmark claim worth checking. Cohere's Hugging Face announcement states that North Mini Code's 33.4 "outperforms Qwen3.5 (35B-A3B)." But Artificial Analysis's independent measurement places it below Qwen3.6 35B-A3B (35.2 vs 33.4) on the same Coding Index. The likely explanation is that Cohere is comparing against an older Qwen generation (3.5) while the current competitive bar is Qwen3.6. The honest framing for developers: North Mini Code is in the same league as the best small open coding models, but it is not the outright leader on raw coding quality. Where it actually wins is elsewhere — speed and license.

The real advantage: speed and total cost of ownership

If you rank North Mini Code purely on intelligence benchmarks, it looks like a solid-but-not-leading small model. That framing misses the point. North Mini Code's genuine edge is throughput, and for high-volume or interactive coding workloads, throughput is what determines your real cost.

In Cohere's internal testing, North Mini Code achieved up to 2.8x higher output throughput than Devstral Small 2 under identical concurrency and hardware — close to three times the work rate. It also showed a 30% advantage in inter-token latency, the metric that governs how smoothly tokens stream during generation. The one place it does not win is time-to-first-token (TTFT), where Devstral Small 2 held a slight edge.

Why this matters more than a two-point benchmark gap: in an agentic coding loop, a model regenerates code, runs tests, reads output, and iterates dozens of times per task. A 2.8x throughput advantage compounds across every iteration, so a model that is marginally less accurate but three times faster often finishes the same task sooner and cheaper. For a developer self-hosting on a fixed GPU budget, throughput-per-dollar is the metric that actually shows up on the invoice — and this is where North Mini Code is built to win.

North Mini Code vs Devstral vs Qwen3.6: which should you use?

The core difference comes down to three competing priorities: raw coding quality, speed, and licensing freedom. Here is the direct comparison for the small open-model class.

North Mini Code vs Devstral Small 2: North Mini Code is roughly 2.8x faster in output throughput and 30% better on inter-token latency, with a comparable size footprint. Devstral keeps a slight TTFT edge and a long track record as a go-to agentic coding model. If you run high-volume or latency-sensitive coding loops, choose North Mini Code. If you want the more battle-tested agentic option today, Devstral is the safer pick.

North Mini Code vs Qwen3.6 35B-A3B: Qwen3.6 leads on the Coding Index (35.2 vs 33.4) and is the stronger model on raw coding quality. North Mini Code counters with the Apache 2.0 license and a speed-tuned profile. If you optimize for benchmark-topping code quality, choose Qwen3.6. If you need a permissive license and maximum throughput on a single GPU, choose North Mini Code.

The honest summary: North Mini Code is not the smartest small coding model, but it may be the most practical one to deploy at scale, because Apache 2.0 plus high throughput plus single-H100 hardware is a combination few competitors match all at once.

Why "sovereign" open-source matters here

Cohere frames North Mini Code around "sovereign AI," and beneath the marketing there is a real, concrete benefit. An Apache 2.0 model you can self-host gives you three things a cloud API cannot: data control, cost predictability, and freedom from vendor changes.

For developers, this translates into specific scenarios where North Mini Code is the right tool regardless of benchmarks. If you work with code that cannot leave your network — regulated industries, proprietary codebases, air-gapped environments — a local Apache 2.0 model is the only viable option, and a frontier cloud model is simply off the table. If you run a coding agent at high volume, owning the weights converts an unpredictable per-token bill into a fixed hardware cost. And because the license is genuinely permissive (not a "source-available" license with commercial restrictions), you can fine-tune it, embed it in a commercial product, and ship it without legal review.

You can deploy it three ways: download the weights from Hugging Face for full local control, run it in a managed environment via Cohere's Model Vault, or call it through the Cohere API or OpenRouter if you want to test before committing hardware.

How to get started with North Mini Code

To start using North Mini Code in under an hour, follow these steps:

1. Test it first via API or OpenRouter before provisioning hardware — call the Cohere API with a few real tasks from your own codebase to check fit.
2. Pair it with OpenCode for the smoothest experience — Cohere specifically trained the model for the OpenCode harness, though it works with most coding agents.
3. For local deployment, download the FP8 weights from Hugging Face and serve them on a single H100; the FP8 quantization is what keeps it inside one GPU.
4. Benchmark on your own tasks, not the leaderboard — measure cost per completed task and tokens-per-second on your workload, since the throughput advantage matters most under your real concurrency.
5. Consult the official documentation for deployment guides, model specs, and cookbooks.

Frequently asked questions

What is North Mini Code?

North Mini Code is Cohere's first open-source agentic coding model, released June 9, 2026. It is a 30B-parameter mixture-of-experts model (3B active) under an Apache 2.0 license, optimized for code generation, agentic software engineering, and terminal tasks, and available for free on Hugging Face.

Is North Mini Code free to use?

Yes. North Mini Code is released under the Apache 2.0 license, which permits free commercial use, modification, and redistribution with no restrictions. You can download the weights at no cost, or pay only for inference if you use the Cohere API or Model Vault.

What hardware do I need to run North Mini Code?

The minimum requirement is a single H100 GPU running the model at FP8 quantization. Because only 3B of its 30B parameters are active per token, it runs at the memory and speed profile of a much smaller model, which is what allows single-GPU deployment.

How does North Mini Code compare to Devstral Small 2?

In Cohere's internal testing, North Mini Code achieved up to 2.8x higher output throughput than Devstral Small 2 on identical hardware and a 30% better inter-token latency, while Devstral Small 2 kept a slight time-to-first-token edge. North Mini Code is the faster model; Devstral remains a strong, established agentic coding option.

Is North Mini Code better than Qwen3.6 for coding?

On the Artificial Analysis Coding Index, Qwen3.6 35B-A3B scores 35.2 versus North Mini Code's 33.4, so Qwen3.6 leads on raw coding quality. North Mini Code's advantages are its Apache 2.0 license and higher throughput. Choose Qwen3.6 for top code quality; choose North Mini Code for a permissive license and speed on a single GPU.

When should I use North Mini Code instead of a cloud model like Claude or GPT?

Use North Mini Code when you need data sovereignty (code that cannot leave your network), predictable fixed costs at high volume, or freedom from vendor lock-in. Use a frontier cloud model when you need maximum reasoning quality on the hardest, lowest-volume tasks where capability matters more than control or cost.

Data sourced from Cohere's official North Mini Code announcement (June 9, 2026), the Cohere Labs Hugging Face model card, Cohere's official documentation, and independent benchmark measurements from Artificial Analysis. Benchmark scores reflect figures published at launch; the Cohere-vs-Artificial-Analysis Coding Index comparison discrepancy is noted in the body. Specifications and availability may change — verify against Cohere's documentation before deploying.