Can you ship Mistral's Codestral weights inside a commercial product?

## What the MNPL actually says The Mistral Non-Production License is a source-available licence written specifically so developers can get hands-on with a model before paying for it. It grants a broad right to use, reproduce, modify and create derivatives of the weights **for development, testing, evaluation and research**, and it withholds the right to use the model or its derivatives in production or in connection with commercial activity. Mistral offers a separate commercial licence for anyone who needs the withheld rights. The practical reading is: your laptop, your CI, your internal benchmark harness, your prototype demo — fine. Anything that serves real users or supports revenue-generating work — not fine, whether the users are external customers or your own engineers. ## Why teams get this wrong Three failure modes recur. **"It's on Hugging Face, so it's open source."** Restricted and permissive weights sit side by side on the hub, download through the same API call, and load with the same code. Nothing in the tooling enforces the licence. **"It's internal, so it isn't commercial."** An internal developer-productivity tool at a for-profit company is commercial activity. "Non-production" is not a synonym for "not customer-facing". **"We fine-tuned it on our own code, so it's ours now."** A fine-tune is a derivative work. So is a merged model, and so is a GGUF or AWQ quantisation someone else uploaded — third-party quantisations frequently carry a sloppy or simply wrong licence tag while the underlying restriction persists. ## The three lawful paths, and how to choose **1. Hosted API.** Calling Codestral through Mistral's platform moves you out of weight-licence territory entirely: hosted use is governed by Mistral's commercial terms of service, which permit production use on a paid plan. This is the intended commercial route and by far the least friction. The tradeoffs are the usual hosted ones — per-token cost, network latency on a workload (code completion) that is exquisitely latency-sensitive, and sending source code to a third party, which is precisely the objection that pushes some teams toward self-hosting in the first place. **2. Negotiated commercial weight licence.** If your constraint is that code must never leave your network — regulated industry, defence, strict IP policy — this is the path that gets Codestral onto your own GPUs lawfully. Budget procurement lead time; this is a contract, not a checkbox. **3. An Apache-2.0 alternative.** Mistral's permissive catalogue includes strong code-capable models — the Mistral Small line and Apache-2.0 coding/agentic releases such as Devstral Small — that you may self-host commercially with no agreement at all. You give up whatever specialist edge Codestral holds on fill-in-the-middle completion quality; often the gap is small enough that the licensing simplicity wins outright. This is the option senior engineers should raise first, because it removes a whole class of ongoing compliance work. ## Running the compliance process For any self-hosted model, not just Codestral: - Record the exact release you deployed and archive its LICENSE file with the deployment, not just a link. - Keep a per-model licence inventory alongside your dependency inventory, and classify each entry as permissive / research-only / non-production / API-only. - Gate on the boundary between environments: a model that is legitimately in your dev cluster may be unlawful in prod, so the promotion pipeline is where the check belongs. - Re-verify on upgrade. Mistral has changed licences between generations, so last year's classification is not evidence about this year's release. - Treat derivatives — fine-tunes, merges, quantisations, distillations — as inheriting the base terms, and record the base model in the artefact's metadata. ## What a strong answer sounds like A weak answer is "Codestral is open, so yes" — it fails on the licence and on the habit of checking. A strong answer names the MNPL, states the specific restriction (non-production, and internal commercial use counts), lists the three lawful paths with their tradeoffs, mentions that derivatives inherit the restriction, and points out that the Apache-2.0 alternative usually costs less total effort than the compliance work of the restricted one.