Three weeks ago, we flagged a problem with Mistral Large 4’s launch: the official announcement listed one price, the live API billed another, and Mistral never explained why. That mystery is closed now — it was a promo, and Mistral said so in writing the whole time, just not anywhere obvious. The bigger news broke on October 29, when Mistral published Large 4’s open weights on Hugging Face, and buried in that announcement is a license that calls itself open while drawing a very specific line around who gets to profit from reselling it.

Quick Summary: What Happened

Detail	Info
Date	October 29, 2026
What shipped	Open weights for Mistral Large 4 (“Le Chonk”), published as Mistral-Large-4-1T-A52B on Hugging Face
Formats	FP8 (54 files) and a quantized NVFP4 build for Blackwell GPUs
License	Mistral Ecosystem Open-Weight License — a modified MIT license
The catch	Operators who resell direct API/gateway access to the model must get a separate commercial license from Mistral
Who’s exempt	End-user products, internal deployments, and services that merely relay requests to someone else’s hosted model
Pricing resolved	The Oct 6 preview’s $0.68/$2.09 rate was a two-week, 50%-off launch promo; full list price is $1.36 input / $4.18 output per million tokens

Bottom line: Open weights, modified-MIT license, one real restriction — and it’s aimed squarely at anyone trying to build a Le Chonk reseller business on Mistral’s own training run.

The Pricing Mystery, Actually Resolved This Time

Quick housekeeping before the license story, because we left readers with an open question and we’d rather close it than let it rot.

Mistral’s own API changelog — a page that existed the entire time, we just hadn’t checked it against the right date — lists the October 6 entry for mistral-large-4 with four words that explain everything: “Launch pricing: 50% off for 2 weeks.” That’s it. No press release, no asterisk on the pricing page, just a changelog line that quietly answered the question everyone was asking in public.

Two weeks after October 6 lands right around October 20, and by the time the open-weight release shipped on October 29, the discount had rolled off. The number that survives is the one we told readers to budget against: $1.36 per million input tokens, $4.18 per million output tokens, with cached input at $0.14. The $0.68/$2.09 figures that Forkast and others reported as simply “the price” were real — just temporary, and apparently disclosed in exactly one place nobody was reading.

We’ll say the same thing we said three weeks ago, because it’s still true: a company pitching itself on European AI sovereignty and enterprise trust should not be relying on a changelog nobody reads to explain a 2x pricing swing. Mystery solved. Communication grade unchanged.

What Actually Landed on Hugging Face

The open-weight drop isn’t identical to the model that previewed on the API in early October. Mistral put the weights through another three weeks of reinforcement learning before publishing them, and it shows in the numbers — Terminal-Bench 4.0 jumped 13 percentage points over the preview score, landing at 41%.

The specs, unchanged from what we covered at launch:

• 1.05 trillion total parameters, ~49 billion active per token (mixture-of-experts)
• Natively multimodal input — text and images — with text-only output
• Trained from scratch on 3,800 Nvidia Grace Blackwell GPUs, entirely in Mistral’s own European datacenters
• Shipped on Hugging Face as Mistral-Large-4-1T-A52B, in two checkpoint formats: FP8 (the main release, 54 files) and a quantized NVFP4 build aimed at Blackwell-class hardware (B200/B300/GB200/GB300)

Mistral is still benchmarking it against the same open-weight set it picked at launch — DeepSeek V4 Pro, Qwen3.8 Max, Kimi K3, and GLM-5.3 — and the post-RL numbers hold up reasonably well inside that group. On Artificial Analysis’s Coding Agent Index, Large 4 now ties GLM-5.3 while using roughly 47% fewer tokens to get there. On the Cyber Index, it leads the open-weight field at 51%. The honest caveat, consistent with what Artificial Analysis found at preview: Large 4 is the strongest Western open-weight model right now. The top Chinese open-weight releases still score higher on Artificial Analysis’s general Intelligence Index. “Best outside the US and China” was always the accurate framing, and it still is.

If you’re trying to place Le Chonk against the rest of the open-weight field, our DeepSeek V4 review and GLM-5 review cover the two models it keeps landing closest to on these benchmarks.

The License Catch: Open, But Not for Everyone’s Business Model

Here’s the part of the announcement that matters more than another benchmark table. Mistral is publishing Large 4 under something it calls the Ecosystem Open-Weight License, which the company and multiple outlets covering the release describe as a modified version of the standard MIT license. Modified how? With one added condition, buried in the part of the license most people skip.

The license’s core restriction kicks in if you operate what it defines as a “Model as a Service” business — providing third parties with API or hosted-endpoint access to the model (or a fine-tuned derivative of it) where those third parties get meaningful control over inputs, parameters, or training data. Trip that definition, and the free ride ends: you need a separate commercial license from Mistral before you can monetize it.

Who Actually Needs to Relicense?

Three categories are explicitly carved out of the Model-as-a-Service definition, meaning they stay under the free, modified-MIT terms:

1. End-user products — an app or feature where the model’s capabilities are embedded inside a specific, fixed harness (think a “summarize this” button), rather than exposed as an open-ended prompt endpoint
2. Internal deployments — running Large 4 inside your own organization, for your own use, with no third party paying for direct access
3. Request relaying — a gateway or router that passes traffic through to a model hosted by someone else, without itself serving the weights

What’s left outside those three buckets, and therefore on the hook for a commercial deal, is narrow but specific: anyone standing up Mistral-Large-4-1T-A52B themselves and selling raw, parameter-controllable API access to it — a direct competitor to Mistral’s own hosted API, built on Mistral’s own open weights. One breakdown of the license put the practical dividing line well: a fixed feature with no exposed prompt stays exempt; a customer-controlled completions endpoint does not.

This is a meaningfully narrower restriction than it might sound at first read. A startup fine-tuning Large 4 for an internal support-ticket classifier is fine. A company embedding it behind a single “draft this email” button in their product is fine. An enterprise running it entirely in-house is fine. The license is really only aimed at one specific business model: becoming an unauthorized Le Chonk hosting provider — the exact role OpenRouter and similar routing services play for dozens of other open-weight models, which is presumably why “mere relaying” got its own explicit exemption.

Why This Matters

Open-weight licensing has quietly become one of the more contested fights in AI, and the terminology gets stretched past its breaking point constantly. “Open weight” has never meant “open source” in the OSI sense — you can download and run the parameters, but the training data, methodology, and now, increasingly, the commercial terms come with strings. Meta’s Llama licenses have carried their own usage restrictions for years. DeepSeek and the other Chinese labs Mistral benchmarks against generally ship under more permissive MIT or Apache terms with no equivalent carve-out.

Mistral’s version is more narrowly targeted than most — it doesn’t restrict fine-tuning, doesn’t cap monthly active users the way some licenses do, and doesn’t require attribution gymnastics. It restricts exactly one thing: becoming a reseller of Mistral’s own inference. That’s a defensible business decision. Mistral spent real capital — 3,800 Grace Blackwell GPUs, a from-scratch training run, its own European datacenters — building this model, and its commercial API is the thing that’s supposed to fund the next one. Giving away the weights while blocking competing hosted resale is a coherent way to get open-weight credibility without torching the revenue line that pays for it.

It’s just not the same thing as “open,” and calling it the “Ecosystem Open-Weight License” rather than something more direct does a bit of work to blur that distinction. Readers who’ve followed how Mistral Forge positions self-hosting and customization, or how the company handled the Voxtral open-source voice release, will recognize the pattern: genuinely useful open releases, wrapped in positioning language that oversells how open they actually are.

Who Should Actually Care About This

Enterprises self-hosting Large 4 for internal use have nothing to worry about. Internal deployment is explicitly exempt, and that’s most of the audience Mistral’s EU-sovereignty pitch was built for in the first place.

Startups building products with Large 4 embedded behind a specific feature are also covered under the end-user product carve-out — as long as you’re not exposing an open completions endpoint that customers can point arbitrary prompts at.

Anyone building or evaluating an inference-hosting business — a GPU cloud, a model-serving platform, an “OpenRouter for Europe” play — is exactly who this license is written for. If your business model is “host open-weight models and sell API access to them,” talk to Mistral’s commercial team before you add Large 4 to the lineup. For context on what self-hosting a 1.05-trillion-parameter model actually costs in hardware, our local LLM guide covers the infrastructure math at a smaller scale, and it scales in the direction you’d expect.

Procurement teams who read our October 7 post can close that open item. Budget against $1.36 input / $4.18 output per million tokens. That’s the confirmed, post-promo number.

The Bigger Picture

Every major AI lab that’s shipped open weights this year has landed on some version of the same compromise: give away the thing that drives adoption and ecosystem goodwill, keep a claim on the thing that generates recurring revenue. Mistral’s version is unusually clean about it — three narrow exemptions, one clear restriction, no ambiguity about what you’re allowed to build. Compare that to the two-week pricing promo that went essentially undocumented in public, and you get a company that writes precise contracts and sloppy announcements. Those are not the same skill, and Mistral is better at the first one than the second.

Our Take

We think the license itself is reasonable, maybe even a model for how open-weight releases should handle the Model-as-a-Service problem — it’s specific, it’s narrow, and it doesn’t pretend to be more permissive than it is once you actually read it. The naming is the part we’d push back on. “Ecosystem Open-Weight License” is marketing language stapled onto a legal document, and it nudges casual readers toward assuming “open” means fewer restrictions than actually exist. Most developers reading a launch announcement are not going to click through to Section 2 of a license file hosted in a model repo.

The pricing resolution is a smaller version of the same habit. The explanation was true and it was public — on a changelog page, as four words, with no link from the announcement that mattered. Mistral keeps shipping genuinely strong technology and then under-communicating the terms attached to it. That gap is going to keep generating posts like this one until someone there decides a sentence on the actual launch page is worth the trouble.

Frequently Asked Questions

What is the Mistral Ecosystem Open-Weight License?

It’s the license Mistral published alongside Large 4’s open weights on October 29, 2026 — a modified version of the standard MIT license. It permits free use, modification, fine-tuning, and most commercial use, with one restriction: operators running a “Model as a Service” business that resells direct API or gateway access to the model must obtain a separate commercial license from Mistral.

Do I need a commercial license to use Mistral Large 4?

For most use cases, no. The license exempts internal deployments, end-user products with the model embedded behind specific features, and services that simply relay requests to a model hosted elsewhere. You only need a commercial agreement if you’re hosting the model yourself and selling third parties raw, parameter-controllable API access to it.

What is Mistral-Large-4-1T-A52B?

It’s the official model name for Mistral Large 4’s open-weight release on Hugging Face — a 1.05-trillion-parameter, 49-billion-active-parameter mixture-of-experts model, published in FP8 and quantized NVFP4 checkpoint formats.

What’s the real price of Mistral Large 4 now?

$1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14. The lower $0.68/$2.09 rate quoted at the October 6 preview was a two-week, 50%-off launch promotion, confirmed in Mistral’s own API changelog, and it has since expired.

How does Mistral Large 4 compare to DeepSeek, Qwen, Kimi, and GLM-5.3 now that it’s open weight?

After an additional three weeks of reinforcement learning past its October preview, Large 4 ties GLM-5.3 on Artificial Analysis’s Coding Agent Index while using about 47% fewer tokens, and leads the open-weight field on the Cyber Index at 51%. It remains the strongest open-weight model built outside the US and China, though top Chinese releases still score higher on Artificial Analysis’s general Intelligence Index.

Can I self-host Mistral Large 4 and sell access to it?

Only with a separate commercial license from Mistral. The Ecosystem Open-Weight License specifically restricts “Model as a Service” businesses — anyone hosting the weights and selling third parties direct, parameter-controllable API access — while leaving internal use, embedded end-user features, and request-relaying services exempt.

---

Last updated: October 30, 2026. Sources: Mistral AI — Mistral Large 4 is now open weight · Hugging Face — Welcome Le Chonk: Mistral Large 4 · Hugging Face — mistralai/Mistral-Large-4-1T-A52B · Mistral AI — API Changelogs · TokenPost — Mistral Releases Large 4 Weights After Three More Weeks of Reinforcement Learning · dev.to — Mistral Large 4’s 1T Weights Are Out · Unite.AI — Benchmark Firm Gives Mistral Large 4 Preview a 38 Intelligence Score.

Related reading: Mistral Large 4 “Le Chonk”: Price Confusion · Mistral Forge Review · Mistral Voxtral TTS: Is Open-Source Voice AI Ready? · DeepSeek V4 Review · GLM-5 Review · Local LLMs Guide