Mistral Deleted the $20M Gate. Read What Replaced It.

Mistral released the weights of Mistral Large 4 (“Le Chonk”) on 29 October 2026 — a 1-trillion-parameter, natively multimodal mixture-of-experts that activates about 49B parameters per token (52B including embeddings and the output layer), with a 1M-token context window — and with the weights came a license that deletes the one number four labs had spent the quarter agreeing on: $20 million. There is no revenue threshold anywhere in the new file. Its single added condition is about what kind of company you are, not how big you are:

“If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose.”

Disclosure: this site runs on Cloudflare Workers and D1 (wrangler.json), does not use Workers AI, and has no affiliate relationship with Mistral.

What the file actually grants

The document is 2,618 bytes, titled the Mistral Ecosystem Open-Weight License, and calls itself “a modified version of the MIT license”. The grant is MIT-shaped: use, copy, modify, merge, publish, distribute, sublicense and sell the weights; “run, deploy, further train or otherwise modify the Model and create derivative works from it”. Then three conditions. Keep the notice. If you or an affiliate operate a Model-as-a-Service business, sign first. And Section 3 takes two classes of use out of that requirement entirely:

“(a) internal use of the Model or its derivative works, defined as any use that does not make the Model or its derivative works available to third parties; or (b) any use of the Model accessed through Mistral’s official products.”

The definition of the gated thing is where the work is done:

“‘Model as a Service’ means giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints), in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data. This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others.”

Put the two carve-outs next to Section 3(b) and the shape is unambiguous. Renting from Mistral is never gated; embedding the model behind a feature of your own product is outside the gate; selling other people access to inference or post-training on these weights is the gate — and the test is whether your customer can steer inputs, parameters or training data, not what you charge them.

The number is gone. The old number is still published

Mistral’s help centre still answers the licensing question with the previous rule, in an article dated 12 August 2026 that is live today:

“Certain models are governed by a modified MIT license, which offers the same permissions as Apache 2.0, with the following exception: Companies with monthly revenue exceeding $20M USD must either: Obtain a commercial license by contacting us, or Use the models via Mistral Studio.”

That sentence describes Mistral’s other modified-MIT models, not Large 4 — and the same vendor’s model overview shows three regimes running side by side:

So “modified MIT” from this vendor now means two different things, and the two Mistral-owned surfaces a deployer is most likely to read disagree about which. The help-centre article is not wrong about the models it covers — it is the wrong document for this one. The card’s own summary is shorter than the file: “for both commercial and non-commercial use — operating a Model-as-a-Service business requires a separate commercial license with Mistral.”

The affiliate clause is wider than the revenue line was

Two phrases in Section 2 carry more than the removed threshold did — our reading of the text, not legal advice and not Mistral’s gloss. The first is “or any of its affiliates”, which the file never defines: on its face, one affiliate operating a MaaS business puts the Licensee’s every commercial use of the weights inside the signature requirement, not only the MaaS use. Meta’s equivalent clause in its Muse Model License, which this site read in full, needed both conditions at once (MaaS and aggregate gross revenue above twenty million dollars over any consecutive twelve months); Mistral’s needs one. The second is “or its derivative works”, which repeats in the definition and in Section 3(a): a fine-tune or distillate you serve to third parties inherits the gate rather than escaping it. The industry’s “open weights” line is moving away from how much you earn and toward whose model your customers are buying — see the same MIT-shaped-grant-plus-one-condition structure in Muse Spark Open, where the badge is due from your first dollar.

Who can sell it today: one, and it is the licensor

As of 30 October, the distribution list is short. OpenRouter has carried a mistralai/mistral-large-4-0 entry since 6 October at $0.70 in / $2.10 out per million tokens ($0.07 cache reads, 1,048,576-token context, 262,144 maximum output) — and its endpoint list for that entry shows exactly one provider, named “Mistral”. The Hugging Face card lists no inference providers at all (gated: false, license tagged other, 278 downloads, 1,240 likes). Cloudflare’s Workers AI catalog still lists only mistral-7b-instruct-v0.2 and mistral-small-3.1-24b-instruct for this vendor. The model APIs of Together, Fireworks and Groq return 401 to an unauthenticated call, so we could not check those three today — treat “no other host yet” as what we could see, not as what is.

Whether OpenRouter’s single-provider listing sits inside carve-out (b) — “mere relaying of requests to models hosted by others” — or inside the gate is a question about its agreement with Mistral, and neither company publishes it.

The date line teaches the same lesson twice

Mistral’s 6 October announcement said the weights were coming, twice: “Weights drop end of this month.” and “We will release the weights by the end of the month.” The Hugging Face repository was created on 30 September, but the history of its main branch begins with one commit dated 29 October 2026 at 14:11:16Z, “Initial release” — and the LICENSE file is present at that commit (fetched by SHA, the same 2,618 bytes), with two further Update README.md commits at 14:28:47Z and 14:51:07Z. For the 29 days between the repo existing and that commit there was no terms text to read at all: no deployment decision, cost estimate or legal review was possible, whatever the calendar said. Mistral’s own news RSS carried no Mistral Large 4 item between 6 and 29 October. Coverage of the window reported an expected 27 October date and 973 people waiting on the card; main shows no placeholder state to compare against, so we could not verify either figure. The release page’s “a full technical report is coming soon” is still a promise — as of today it is not published.

If you are the one deploying

1. Renting it: Section 3(b) puts you outside the gate, and the sticker is public — $0.70 / $2.10, whether through Mistral or an aggregator relaying Mistral. Choose on cache hit rate and latency, not on who is allowed to sell it.
2. Shipping a feature that uses it: carve-out (a) is on your side while the model stays inside your product. The moment a customer can steer the parameters, run post-training on their data, or point their own traffic at the weights, you have moved toward the centre of the definition.
3. Reselling inference: size no longer buys you anything. A two-person MaaS needs a signed agreement before its first commercial request; a $200M product company embedding the model may need none. The published path is Mistral’s contact form; no price for that agreement is published anywhere we could read, so budget the negotiation, not just the tokens.
4. Fine-tuning or distilling: derivatives are named in Section 2 and in the definition. If you serve the fine-tune to third parties, the gate travels with it.
5. The habit worth keeping: a vendor’s help-centre answer is not the license. Read the LICENSE in the model card of the specific model — and on Hugging Face it is tagged other, so you cannot filter your way to it.