Europe’s best-funded AI lab just released its biggest model, and the first thing to know is that you cannot actually run it yet. On Tuesday, October 6, Mistral AI launched Mistral Large 4, a multimodal model with roughly 1 trillion parameters. The company’s own nickname for it is “Le Chonk.” The weights, which are what make a model “open,” are not out. Mistral says they will arrive in about three weeks, once safety testing is finished.
The short version
- Mistral Large 4 (ML4) has about 1 trillion parameters and handles text and images
- It is available now only through a public guardrail endpoint, not as a download
- Open weights are promised in about three weeks, after safety testing
- It was trained on just 4,000 Nvidia GPUs, which Mistral says is two to three times fewer than Chinese rivals use
- Benchmark results are still pending, so the “best in class” claim is a goal, not a fact
What Mistral actually released
Mistral Large 4 is the successor to Mistral Large 3, which shipped in December 2025 as an open-weight mixture-of-experts model with 675 billion total parameters and 41 billion active per token. ML4 pushes the total to about 1 trillion. That is a jump of roughly 48% in size in under a year.
The company is pitching it as a model that can beat both American closed models and Chinese open ones. Mistral VP of Science Pierre Stock told TechCrunch that the goal is to be best in class among open-weight models, especially outside China. He also said focused training could let ML4 beat closed models in specific areas that matter to Mistral’s customers.
Why the weights are delayed
This is the most unusual part of the launch. Mistral built its reputation on handing models to the public. This time it is holding the weights back while it runs safety testing. Stock said the company will work with trusted partners and governments in the meantime, so the open weights “can be used to defend, but not to perform malicious attacks.”
The reason is security. Concern has been growing among Mistral’s core customers, which are enterprises and institutions. A trillion-parameter model with strong cybersecurity skills is also a model that could be misused. At the same time, Stock argued that open weights are easier to audit than closed ones, which is a point enterprise buyers care about. So Mistral is trying to have it both ways: ship the access now, ship the files once it is comfortable.
The efficiency claim worth watching
The detail that jumped out at us is the compute. Mistral says ML4 was trained entirely on its own infrastructure, using only 4,000 Nvidia GPUs. Stock described that as two to three times fewer than its Chinese competitors and significantly fewer than closed-source labs use.
| Detail | Mistral Large 3 | Mistral Large 4 |
|---|---|---|
| Release | December 2025 | October 6, 2026 |
| Total parameters | 675 billion | About 1 trillion |
| Open weights | Yes, at launch | Promised in about 3 weeks |
| Training hardware | Not covered here | 4,000 Nvidia GPUs, Mistral’s own compute |
| Focus areas | General purpose | Cybersecurity, finance, chip design |
If that number holds up, it matters because compute is the main cost of staying at the frontier. We saw the other side of this recently, when DeepSeek closed the gap with the best US model to about 3% using cheaper, open models. Mistral is making a similar argument from Europe: you do not need the biggest cluster on earth to get close to the top.
One caution. The figure comes from Mistral itself, and no outside lab has verified it. Training efficiency is also easy to define in flattering ways, for example by leaving out earlier experiments or data preparation. Treat it as a claim until the technical report lands.
Why chip design and cybersecurity
The target use cases are not random. Stock named cybersecurity and finance as optimized areas, and also chip design. That last one lines up neatly with Mistral’s investors. Dutch chip equipment giant ASML led its Series C, and Samsung led its Series D last month at a valuation of €21 billion, about $24.39 billion. We covered that round in our piece on Mistral’s record European funding and its sovereignty pitch.
A multimodal model that can read diagrams, layouts and documents is a natural fit for semiconductor work. For Samsung and ASML, a European model they helped fund and can run on their own hardware is more appealing than depending on a US API that could change terms or be switched off.
What to keep in perspective
- No benchmarks yet. Mistral hopes ML4 will be best in class among open-weight models, but it has not shown the numbers
- “Open” is not here yet. The weights are a promise, and safety testing could take longer than three weeks
- Size is not quality. A larger parameter count helps, but training data and post-training decide how useful a model is
- Self-reported efficiency. The 4,000 GPU figure has not been independently checked
The “third way” and what it means for you
French president Emmanuel Macron has described Mistral’s approach as “a third way in AI.” The idea is simple. The market is splitting into closed models that a vendor can unplug, and open models that are often made in China. Mistral wants to be the option for buyers who want neither, especially governments and regulated industries in Europe.
- If you build products, wait for the weights and the benchmarks before switching anything. The endpoint is the only way to test it today
- If you run an enterprise, the self-hosting promise is the real draw, particularly for finance and security teams with strict data rules
- If you follow the AI race, watch the leaderboards in three weeks. That is when the claim of beating open and closed rivals gets tested for real
For now, Le Chonk is an announcement with an asterisk. If the weights arrive on time and the scores hold up, Europe has a credible frontier contender. If they slip, it will look like a very large model that mostly lives behind a fence.
Sources and further reading
- TechCrunch: Mistral’s new 1T model aims to leapfrog closed and open rivals
- IntuitionLabs: Mistral Large 3, an open-source MoE LLM explained
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.

