A blogger reads an announcement: "One trillion parameters." Quick math. One trillion. That's a 1 followed by twelve zeros. Then comes the second sentence: "Weights to follow."
Welcome to Mistral Large 4.
What happened?
French AI company Mistral released a preview of its biggest model yet in early October. Reports on it rely partly on aggregator sites rather than detailed vendor documentation, so details may still change.
According to current reports, it uses a Mixture-of-Experts architecture (a model made of many specialists, where only a small part works on each request, explained here). The model is huge, but only uses a fraction of it per request. It also has a very large context window, meaning it can handle lots of text at once.
Open? Not yet.
The interesting part: Mistral has announced open weights. Weights are the learned numbers of a model. Whoever has them can run the model on their own hardware. More in our article on open AI models.
As things stand, they are not available for download. The end of the month is mentioned, and the license terms are open. And "open" does not mean "free to use": the license decides whether you may use the model commercially.
Honestly: a model this size will not run on your laptop. Nor on your gaming PC. Open weights help companies, researchers and hosting providers first. You benefit indirectly, through more choice and price competition.
What does it cost?
Through the API (the interface for developers) the model costs money, billed per token. A token is roughly a syllable. Reports currently mention an introductory price that is meant to rise later. The numbers partly contradict each other, so I won't quote any. Check Mistral's pricing page before you calculate.
Whether Large 4 also powers Le Chat has not been confirmed by Mistral, as far as we found. Le Chat is the chat service we covered here. It is free to try, but with limits. As with every provider: heavy users end up on a paid plan eventually.
Is it for me?
In short:
- Beginners: Watching is enough. Try Le Chat if you haven't yet. You don't need to switch anything.
- Developers: Test the model once access is stable. Comparing it with other large open models like Qwen is worthwhile.
- Privacy fans: A European provider is a plus, but it doesn't replace reading the contract terms.
Mistral claims to keep up with leading Chinese models. That is a vendor claim. Independent tests take time.
Bottom line
Europe has another big model in the race. That's good news. Whether it keeps its promise will show once the weights are out and others measure it.
Until then: stay curious. Not every number with twelve zeros is a breakthrough.