Mistral just dropped its biggest model ever, and the AI world is calling it Le Chonk. Mistral Large 4 is a trillion-parameter monster that runs like a much smaller model, and you can try it today without paying for ChatGPT or Claude.
If you have been watching the open-weight AI model space, you know the pattern by now. The best free models keep coming out of China, while US and European labs lock everything behind subscriptions and APIs. Mistral wants to end that streak. Its new flagship, Mistral Large 4, is the company’s claim to the best open-weight model built outside China, and the numbers back up at least part of that story.
What Mistral Large 4 actually is
Le Chonk is the unofficial nickname, but the official name is Mistral Large 4. The French company launched a public preview on October 6, and the full model weights follow on October 27. Until then, you can use it through the preview API on Mistral Studio, and the endpoint is called mistral-large-4-preview.
The spec sheet looks wild on paper. One trillion total parameters sounds like something only a datacenter could run, but here is the trick: Mistral Large 4 uses a sparse Mixture of Experts architecture. Only 49 billion of those parameters activate for any given task. In practice, that means runtime compute looks more like a 49B model than a 1T monolith, which keeps it fast and keeps the price down.
It is also natively multimodal. The model handles both text and images out of the box, no adapters required.
What it costs to try Le Chonk today
The preview API is live at $1.36 per million input tokens and $4.18 per million output tokens. For context, that input price undercuts both GPT-6-Astra and Claude Opus 5.5, and the output price is competitive with the big closed models.
Access is simple. Create a Mistral Studio account, pick the preview endpoint, and start sending requests. No waitlist, no enterprise contract.
How Le Chonk performs against other models
Benchmarks are where things get interesting, and where the marketing needs a little unpacking. Mistral reports strong results in a few specific areas rather than a clean sweep across the board.
On DeepSWE, an agentic coding benchmark built on real software engineering tasks, Mistral reports a score around 62 percent. That lands it just ahead of Zhipu’s GLM-5.3 and clearly in front of the other open-weight rivals.
| Model | DeepSWE score | Weight access |
|---|---|---|
| Mistral Large 4 | 62% | Open (Oct 27) |
| GLM-5.3 | 61% | Open |
| DeepSeek-V4-Pro | 57% | Open |
| Qwen 3.8 Max | 51% | Open |
| Reflection Beam | 44% | Open |
Cybersecurity is the surprise standout. On a benchmark that asks models to reproduce and patch a real vulnerability, Mistral Large 4 scored 82 percent, the highest number reported so far. Several leading closed models score near zero on the same test. Not because they cannot, but because they refuse the task on safety grounds. Mistral also claims 93 percent on Cybench and a top-five spot on the Artificial Analysis Cyber Index.
Visual grounding is another win. Mistral says Mistral Large 4 beats even closed frontier models there, meaning it is better at pointing at exactly where things are inside an image.
Why should you care about that last one? Because visual grounding is the skill behind automation that clicks buttons and reads screenshots. If you build agents that fill forms, pull numbers from invoices, or navigate dashboards, that single benchmark line matters more than any leaderboard position.
Now for the honest part. On the general Artificial Analysis Intelligence Index, Le Chonk scores 38, tied with GPT-6 Luna. Claude Opus 5.5 leads that chart at 58. The best Chinese open models still score higher too. So “best open-weight model outside China” is accurate, and “best model, period” is not. As a result, temper expectations if you plan to throw frontier-level reasoning problems at it.
Why open weights even matter here
You might wonder why anyone cares about open weights when closed APIs are everywhere. Three reasons, and they all hit your wallet or your privacy.
First, control. When the weights drop on October 27, you can download the model, around 240GB of it, and run it on your own hardware. Nobody can rate-limit you, deprecate your workflow, or change the terms overnight. Companies got burned by exactly this in 2026, which is why Hugging Face ended up relying on a Chinese model for its own stack.
Second, privacy. Sensitive data never has to leave your infrastructure. For law firms, hospitals, and anyone handling documents under NDA, that is the whole ballgame.
Third, cost at scale. The API price is good, but self-hosting gets cheaper than any API once your volume is high enough. Quantized versions should follow within days of the release, and those cut the hardware bill further. We covered the tradeoffs in our open vs closed AI models guide if you want the full breakdown.
The catch: running it yourself is not cheap
Honesty time. Self-hosting Le Chonk at full size takes serious iron, roughly eight A100 80GB cards or an equivalent H100 setup. That is a far cry from the gaming-PC setups we usually talk about, like when we ran Qwen 3.8 locally on a gaming PC.
The realistic path for most people looks like this: use the cheap preview API now, then check the quantized community builds that show up after October 27. Those versions trade a little quality for a much lower hardware bar, and honestly, for everyday tasks the difference is often hard to notice.
There is also a three-week red-teaming window before the weights release. Cybersecurity firms, vetted partners, and government agencies get expanded access during that period. This is Mistral testing the model in the wild before anyone can download it permanently.
Should you switch from ChatGPT or Claude?
Switching is the wrong frame. Think of Mistral Large 4 as a second tool in the drawer.
For coding agents and automation work, especially anything security-adjacent, it is a genuinely strong pick, and the open-weights release makes it the only top open model you can fully own outside Chinese labs. For deep general reasoning and writing polish, Claude Opus 5.5 still leads the index, so keep your subscription if that is your daily driver. We compared the open and closed tradeoffs in our GLM 5.3 open-weight breakdown as well.
My take? Start with the preview API while it costs pennies, test it on your actual workload, and decide about self-hosting when the weights land. If your projects involve security tooling or document-heavy automation, I would test this one first.
Your move this week
The window between now and October 27 is basically a free trial of the weights that are coming. Set up a Mistral Studio account, run your three most common AI tasks through mistral-large-4-preview, and compare the output quality against whatever you use today. If it keeps up, you will be ready to self-host or negotiate from a position of strength. If it doesn’t, you have lost an afternoon and almost nothing else.
