Mistral Large 4: a 1-trillion-parameter European model with open weights on the way
On October 6, Mistral introduced Mistral Large 4 (nicknamed "Le Chonk"), its largest model, trained from scratch in its own European data centers.
What's announced
- Mixture-of-experts architecture: about 1 trillion total parameters, roughly 49 billion active per request.
- Multimodal (image input) and over 160 languages, including every official EU language.
- Trained on about 4,000 Nvidia Grace Blackwell GPUs over two months.
- Mistral calls it the strongest open-weight model developed outside China. That's the vendor's claim, on aggregate benchmarks: wait for independent verification.
Access and pricing
| | Detail | |---|---| | API | limited preview on Mistral Studio | | List price | $1.36 input / $4.18 output per million tokens | | Preview price | half that ($0.68 / $2.09) | | Weights | release announced for October 27 |
The thing to watch: the license
Large 3 was Apache 2.0. For Large 4, press reports a custom Mistral license whose terms haven't been published. "Open weights" doesn't mean "open source": before building a product on it, read the terms on commercial use, redistribution and fine-tuning.
What it means for self-hosting
A ~1T-parameter model, even with ~49B active, is out of reach for a VPS or a single GPU: every expert must sit in memory. For most teams the value will be the European API (data hosted in Europe, a sovereignty argument) rather than local hosting. Smaller variants, if any, will likely matter more for homelabs.
Bottom line: competitive API pricing, but wait until October 27 to judge the license, the weights and independent results.
Sources: TechCrunch, SiliconANGLE, Quartz, TechXplore, Beam AI.