What if one of the most capable AI models around was just sitting there, free to download? That’s more or less what happened on October 29. French AI lab Mistral released the open weights for Mistral Large 4, a model with about one trillion parameters.
The team’s nickname for it is “Le Chonk,” which tells you something about the size and the company’s sense of humor. But the bigger story isn’t the nickname. It’s about who gets to control powerful AI, and if you really have to rent it from someone else.
So what did Mistral actually release?
Mistral first showed Mistral Large 4 as a public preview earlier in October. About three weeks later, it put the full weights on Hugging Face for anyone to grab, as the company explains in its release post.
Here’s the quick spec sheet. The model has roughly a trillion parameters in total, but only 52 billion switch on for any given task. That’s because it’s a “mixture of experts” model. Think of it as a huge team of specialists where only the right few show up for each job.
It reads text and images, then writes text and code. If you’re fuzzy on how that works, here’s our plain guide to multimodal AI. Mistral says it handles more than 160 languages too.
Mistral also trained it from scratch in Europe, on 3,800 Nvidia Grace Blackwell GPUs, according to The Next Web. Since the preview, the team spent about three more weeks on reinforcement learning (basically, training by trial and reward). The focus was coding and defensive cybersecurity.
Why open weights are a bigger deal than they sound
Most big-name AI models keep their weights private. You use them through an app or an API, and the company sets the price, the rules and your access. We saw that debate play out again when Meta kept its most powerful model closed.
Open weights flip that. Weights are basically the model’s brain: the huge set of numbers it picked up during training. If you have them, you can run the model on your own servers, tweak it with your own data, and keep everything in-house.
“AI is going to be like electricity,” Mistral’s VP of Science, Pierre Stock, told The Next Web. In his view, no company should depend on a provider that can cut its access “from one day to another”. And once weights are out there, nobody can take them back.
Stock also pointed out that Chinese labs dominate open models right now. He argues they don’t guarantee they’ll keep publishing weights. So Mistral is pitching itself as the open option Europe, and frankly everyone else, can count on.
How good is Mistral Large 4, really?
On Mistral’s own numbers, pretty good. Take Terminal-Bench 4.0, a test of “agentic” coding, where the AI works through multi-step tasks in a terminal by itself. The score jumped from 28% in the preview to 41%. That edges out GLM-5.3 at 40%, and it’s well ahead of DeepSeek V4.1 Flash at 31% and Kimi K3 at 21%.
It also got more efficient. Mistral says it hit that higher score while using 30% fewer output tokens. On a coding index from Artificial Analysis, it roughly matches GLM-5.3 while using 47% fewer tokens per task. Fewer tokens means a smaller bill, so that part matters.
There’s a safety angle too. Mistral says the model refused malicious cyber requests 96.6% of the time on average across three public test sets. The company also says it’s updating its guardrails after spotting signs of cyber misuse in live traffic.
Now the honest caveat. Almost all of these numbers come from Mistral itself. Independent testers haven’t confirmed them yet, so treat them as a promising first look instead of the final word.
The catch: you’ll need serious hardware
“Free to download” doesn’t mean “runs on your laptop.” According to The Next Web, even a quantised version (a compressed copy that trades a bit of precision for a smaller size) needs four Nvidia B200 or B300 GPUs. That’s data center gear, not a gaming PC.
Mistral does plan smaller, distilled versions in the coming months, including one that can run on a laptop. It hasn’t committed to dates, though.
Then there’s the license. Mistral calls it the Mistral Ecosystem Open-Weight license, and it’s a tweaked version of the MIT license. You can use, modify, fine-tune and sell products on top of it for free. But if your business sells the model itself as a hosted service, you’ll need a separate commercial license from Mistral.
Don’t want to run it yourself? It’s also on Mistral’s API and on platforms like OpenRouter and Microsoft Foundry. Mistral’s page lists $0.70 per million input tokens and $2.10 per million output tokens. For comparison, OpenAI’s GPT-6.1 Terra lists $1 and $5 for the same amounts.
Why smaller countries should watch Mistral Large 4
Here’s where it gets interesting for places like Armenia. Banks, government agencies and startups often handle data they’d rather not send abroad. An open model you host yourself keeps that data at home.
ENISA, the EU’s cybersecurity agency, backed that idea in Mistral’s announcement. Open-weight models, it said, “allow organisations to keep control of the technology their resilience depends on”. A French bank, La Banque Postale, went further and said “technological sovereignty is not optional”.
We think that’s the real headline. For smaller markets, the choice used to be simple: use a big US model or go without. Now there’s a serious third option, even if running it still takes real money and skill.
Also, keep an eye on those planned laptop-sized versions. If they deliver, a small team in Yerevan could test a capable open model without renting a data center first.
So, should you care about Mistral Large 4?
If you just chat with AI on your phone, Mistral Large 4 won’t change your week. But if you build software, work with sensitive data or care about who controls the tools we’re all starting to rely on, it deserves a look.
Here’s a simple way to start. First, try it through the API on a task you already do with another model. Then compare the quality and the bill. If it holds up, you’ll know whether self-hosting is worth the hardware later.
Our take: open models are closing the gap fast, and that’s good news for everyone outside the big AI hubs. The question now is simple. Will the big closed labs feel enough pressure to open up too, or will they double down on keeping their best models private?