Mistral Large 4 is a 1-trillion-parameter AI model that Mistral opened in public preview on October 6, 2026. You can call it through the API today, but the open weights you’d need to run it yourself aren’t out yet. The company says they’re coming by the end of October.
That gap matters. Here’s what you can actually do with Mistral Large 4 right now, and what to hold off on.
The short version, before the details:
- Announced: October 6, 2026, as a public preview
- Size: 1 trillion total parameters, 49 billion active at a time
- Access today: Mistral’s moderated API only
- Open weights: promised for the end of October, license not yet published
- Trained on: roughly 4,000 Nvidia Grace Blackwell GPUs in European data centers
If you only want to try it, the API is enough. If you want to host it yourself, you’re waiting.
What Mistral Large 4 actually is
It’s a mixture-of-experts model. That means only a slice of the network, 49 billion of the 1 trillion parameters, does the work on any given token. You get big-model knowledge at closer to mid-size running costs.
It’s also natively multimodal, so it handles text and images in one model rather than bolting vision on afterward. Mistral says it was trained on more than 160 languages, including every official EU language.
Mistral has nicknamed it “Le Chonk,” which tells you how the team feels about the size. The practical point is that this is Mistral’s biggest model yet, and it’s aimed at the same crowd that has been watching open models like Reflection AI’s Beam chase the closed labs.
Why you can’t download it yet
Mistral is staging the release on purpose. Right now the model sits behind a moderated API while the company red-teams it with cybersecurity firms, vetted partners, and state authorities.
According to Runtime Wire, those red-team partners get a version with reduced moderation and expanded cyber capabilities. Mistral’s pitch is that security teams should be able to work under their own policies, including on vulnerability research that other providers’ safeguards might block.
There’s a trade-off here. The same report notes that once weights are public, safeguards are easier to strip out. Testing first is a sensible move, but it doesn’t remove that risk.
If you’ve followed the Mistral Vibe vulnerability story, you know security is already a live topic for this company.
How much Mistral Large 4 costs to try
Pricing is where reports differ, so read this part carefully. The independent site Kingy.ai says Mistral’s model card shows $0.68 per million input tokens and $2.09 per million output tokens. Those are half of the crossed-out standard rates, which would be $1.36 and $4.18.
Kingy.ai also says it couldn’t confirm how long the discount lasts. Treat the lower numbers as a launch price, not a promise.
| Token type | Preview price (per 1M) | Standard rate shown |
|---|---|---|
| Input | $0.68 | $1.36 |
| Cached input | $0.07 | $0.14 |
| Output | $2.09 | $4.18 |
Kingy.ai adds that regional EU inference carries a 10% uplift. If you’re comparing it with other budget models, our look at DeepSeek V4.1 Flash pricing is a useful yardstick.
The benchmark numbers, and how far to trust them
Mistral and the outlets covering the launch cite strong security results. The company-reported scores include 82% on a vulnerability reproduction and patching test and 93% on Cybench, a set of 40 security exercises. Prompt-injection resistance is listed at 93.3%.
On general work, the picture is more modest. Reported figures include 49.8 on a coding agent index and 59.9% on AutomationBench. Those are Mistral’s own numbers, so wait for independent testing before betting a project on them.
The context window is the clearest warning sign. Mistral advertises 1 million tokens, but Kingy.ai says Vals lists about 512K and Artificial Analysis about 524K. Test your actual endpoint with a long document before you design around the bigger number.
Who should pay attention to Mistral Large 4
Developers in Europe have the clearest reason to watch. The model was trained in European data centers, and Mistral offers regional EU inference, which can simplify data-residency questions for some teams.
Security researchers are the other obvious audience. The reduced-moderation red-team version is clearly built for them, though you’ll need to be a vetted partner to get it.
Everyone else can simply wait. A new open-weight giant usually spawns smaller, cheaper variants within weeks, and those are the ones most people end up using.
Frequently Asked Questions
Is Mistral Large 4 open source?
Not yet. Mistral says open weights arrive by the end of October 2026, but the license terms haven’t been published. Check commercial-use rules before you build on it.
Can I run Mistral Large 4 on my own computer?
Almost certainly not on a normal PC. Even with only 49 billion active parameters, the full 1-trillion-parameter model needs serious server hardware to hold in memory. Smaller, trimmed versions may show up later.
How do I try it today?
Use Mistral’s public preview through its API. It’s a moderated endpoint, so some security-related requests may be blocked.
Is the preview price permanent?
No one has said so. Independent reporting found no stated end date or guarantee for the 50% launch discount.
Our take
Try the API if you’re curious, especially for multilingual or security-flavored work, but don’t commit a product to it yet. The license, the real context limit, and the final price are all still open questions. Check back after the weights land at the end of October, when independent benchmarks should settle most of them.
Leave a Reply