Which Mistral models are Apache-2.0 and which need a commercial licence?

## Why Mistral has two licence tracks Mistral's commercial strategy is deliberately split. It publishes genuinely open weights to build adoption and community mindshare, and it keeps its strongest and most commercially valuable models either behind the API or behind a paid weight licence. Understanding which track a given model sits on is the first question in any "can we self-host Mistral?" conversation, and it is what interviewers are actually probing — not the parameter counts. ## Track 1 — Apache 2.0 Apache License 2.0 is a permissive OSI-approved licence. For model weights it practically means: you may download them, run them anywhere including in production, modify and fine-tune them, redistribute the originals or your derivatives, embed them in a closed-source commercial product, and you owe no royalty. There is **no monthly-active-user cap, no revenue threshold, and no obligation to open-source your derivative**. The obligations are light: retain the copyright and licence notices, state significant changes, and note that the licence includes an express patent grant that terminates if you sue over patents in the work. As of mid-2026 the Apache-2.0 side of the Mistral catalogue includes Mistral 7B, Mixtral 8x7B and Mixtral 8x22B (the sparse mixture-of-experts pair), Mistral NeMo 12B, Pixtral 12B (vision-language), and the Mistral Small line from Small 3 onward — the 24B-class model that in practice replaced Mixtral 8x7B for most self-hosting teams. Several agentic/coding and audio releases have also landed under Apache 2.0. ## Track 2 — the Mistral Research License The MRL is a source-available, **non-commercial** licence. Mistral publishes the weights so researchers and engineers can study, benchmark and experiment with them, but any use "in connection with a commercial activity" — internal or external — requires a separate Mistral Commercial License obtained from the company. Mistral Large 2 and Pixtral Large are the canonical MRL releases. Two consequences people miss: internal business use inside your company still counts as commercial, and a fine-tune or distillation of MRL weights is a derivative that inherits the restriction. ## Track 3 — the Mistral Non-Production License The MNPL, used for Codestral, is narrower still. It allows development, testing, research and evaluation, and explicitly excludes production and any commercial use. It exists so a developer can prototype with the model locally before deciding to pay for either a commercial weight licence or hosted API access. ## Track 4 — no weights at all Mistral Medium is API-only: there is no public weight release, so the only lawful path is la Plateforme (or a cloud partner such as Azure AI Foundry, AWS Bedrock or Google Vertex AI where Mistral models are resold). API use is governed by Mistral's terms of service, not by a weight licence at all — a different legal instrument covering data handling, rate limits and acceptable use. ## The hazard: availability is not permission The single most common mistake is treating a Hugging Face model page as proof of open source. Mistral publishes MRL and MNPL weights on the same hub, in the same format, downloadable with the same `from_pretrained` call. Nothing in the tooling stops you; only the LICENSE file does. Equally, tags like "open-weights" or "open model" in blog posts do not mean OSI-open. ## Practical compliance workflow 1. For every Mistral model you intend to ship, record the exact release name and read that release's LICENSE and any accompanying use policy. 2. Classify it: Apache 2.0 (ship freely), MRL (research only — needs a commercial agreement), MNPL (non-production only), or API-only. 3. Track derivatives: fine-tunes, merges, quantised GGUF builds and distilled students inherit the base licence, and third-party quantisations on the hub often carry a different, sometimes wrong, licence tag than the original. 4. Re-check on upgrade. Mistral has changed licences between generations — Mistral Small moved to Apache 2.0 at version 3 — so the answer for one release is not automatically the answer for the next. ## What the answer signals in an interview A candidate who can name the three licence instruments and state the practical difference (free commercial self-hosting vs research-only vs non-production) is showing that they can run a vendor-selection conversation with legal in the room. A candidate who says "Mistral is the open-source one" is showing they have only read the marketing.