Picture this: a small team of developers in a cramped office, staring at a cloud bill that has quietly doubled in a month. Their app runs on a rented AI model, and every price rise lands on them with no warning. Then, one morning, a download link for Mistral Large 4 appears that changes the maths entirely.
That is the promise of Mistral Large 4, the new flagship model from Mistral AI. It is now generally available, and unlike the biggest models from its American rivals, you can download its weights and run it yourself.
Mistral Large 4 is a 1.05-trillion-parameter model, built as a granular mixture-of-experts system with 52 billion parameters active on any given token. It handles text and images, and it reads up to one million tokens of context in a single pass.
The model first appeared in public preview on October 6, 2026. Then, almost without ceremony, its status flipped. Mistral’s documentation page now dates the model’s general availability to November 5, 2026, and the change was first spotted by independent trackers on November 6.
Mistral Large 4 Goes Generally Available — Quietly
There was no keynote, no countdown and no press conference. Mistral simply updated the model’s page on its documentation site, changing the status line from “Public Preview” to “GA” under the Mistral Ecosystem Open Weight License v26.10.
That low-key arrival fits Mistral’s style. The Paris-based company has always preferred shipping over showmanship, and the quiet flip to general availability says the model was ready before the marketing caught up.
The timing matters for a different reason, though. The first week of November was crowded with frontier launches, including OpenAI’s GPT-6.2 family and Google’s Gemini 4 Carbon, which we covered here on Wasl Tech. A model that arrives without a launch event can easily be missed in that noise.
It should not be. Mistral Large 4 is the largest model Mistral has ever made available to download, and its arrival reshapes the open-weight end of the market in a single stroke.
The Weights Are Free to Download
The most important fact about Mistral Large 4 is not its size. It is that the weights are downloadable. They have been available since October 29, 2026, under the Mistral Ecosystem Open-Weight License v26.10.
That makes it the biggest Mistral model anyone can take home. You can run it on your own hardware, fine-tune it for your own work, and keep your data inside your own walls. No other frontier-class model from a European lab offers that today.
The licence does come with a catch worth reading. Commercial use is permitted, but companies that want to offer the model as a hosted service need a separate commercial licence from Mistral. The weights are open; the right to rent them out to others is not automatic.
That distinction matters for cloud providers and API resellers. For everyone else — researchers, startups, and companies running their own infrastructure — the download is the whole point.
What Mistral Large 4 Costs on the API
For teams that would rather rent than run, Mistral Large 4 is priced at $0.70 per million input tokens, $0.07 for cached input, and $2.10 per million output tokens on the Mistral AI Studio API.
That price is itself a story. When the model launched in preview on October 6, it cost $1.36 in and $4.18 out. On October 30, Mistral cut both figures by roughly half, according to AI Signal, which tracks the company’s pricing page.
Mistral’s changelog describes the current price as launch pricing at 50 percent off for two weeks, and it has said nothing about what happens after that window. The price could rise again, so the $0.70 figure should be treated as an opening offer rather than a permanent fixture.
Even so, a concrete example shows what the price means in practice. A request of 10,000 input tokens and 1,000 output tokens costs $0.0091 on Mistral Large 4 — less than a cent, according to AI Signal’s calculations.
Two Older Models Retire to Make Room
Mistral Large 4 does not arrive alone. It replaces two models at once, and both were deprecated on October 31, 2026: Mistral Medium 3.5 and Mistral Large 3.
Medium 3.5 was a 128-billion-parameter dense model released on April 28, 2026, under a modified MIT licence. It cost $1.50 per million input tokens and $7.50 for output. Moving to Large 4 cuts its output price by more than three and a half times.
Large 3, released on December 2, 2025, was a different story. It was the largest Apache 2.0 model Mistral had ever shipped, at 41 billion active parameters out of 675 billion in total. It was also cheaper, at $0.50 in and $1.50 out.
That creates an odd asymmetry. Teams moving from Medium 3.5 to Large 4 pay less on every line of the bill. Teams moving from Large 3 pay about 40 percent more, whatever the mix of input and output tokens.
What they get in return is a much longer memory. Both retired models topped out at a 256,000-token context window. Mistral Large 4 reads one million tokens — four times as much — which means whole codebases and long documents fit in a single request.
Both deprecated models remain accessible for now. Mistral’s documentation lists April 30, 2027, as their retirement date, after which requests to their identifiers will fail. There is one trap to watch: the alias mistral-large-latest still points at the deprecated Large 3, so anyone using it is still running the old model at the old price.
Why Mistral Large 4 Matters for Open-Weight AI
The open-weight race has been heating up for months. Meituan’s LongCat-2.5 arrived with 1.6 trillion parameters under the permissive MIT licence, while Alibaba’s Qwen4-Max is still preview-only, with no downloadable weights at all.
Mistral Large 4 sits between those two poles. It is not as permissively licensed as LongCat-2.5, and it is not as large. But it is a genuine flagship from a European lab, with weights you can download today and a one-million-token window that the MIT-licensed rival does not match.
That combination gives European companies something they have lacked: a frontier-class model they can run on their own infrastructure, under European law, without depending on an American or Chinese provider. For banks, hospitals and governments with strict data rules, that is not a nice extra. It is the whole decision.
The contrast with the closed frontier is just as sharp. Google’s Gemini 4 Carbon and OpenAI’s GPT-6.2 Astra are more powerful on paper, but they only exist behind an API. Mistral Large 4 trades a little raw capability for something neither of them offers: ownership.
What Happens Next for Mistral Large 4
The first test will be the benchmarks. Mistral has not attached headline benchmark scores to the general-availability listing, so independent evaluations over the coming weeks will show where the model really stands against its closed rivals.
The second test is the price. If the launch discount expires and the $0.70 rate rises, the value calculation changes — especially for teams migrating from the cheaper Large 3. Watch Mistral’s pricing page, not its press releases, for that signal.
The third test is adoption. Downloadable weights only matter if people actually download them. The model’s size — over a trillion parameters in total — means running it needs serious hardware, so the first users will be well-funded teams and cloud providers rather than hobbyists.
Still, the direction is clear. A year ago, the idea of downloading a trillion-parameter model from a European lab sounded like a fantasy. Today, Mistral Large 4 makes it a download link. The rest of the industry will have to answer that.



