Mistral says it trained ML4 “from scratch” using 3,800 Nvidia Grace Blackwell GPUs in its data centers in Europe and much of its training data was multilingual
Le Chonk — Today, we're launching a public preview of Mistral Large 4. Unofficially ML4, very officially: le Chonk.
Context & Ripple Effects
Mistral’s flagship line has moved well beyond the 123B-parameter Large 2, whose commercial use required a separate license in 2024. Its Small 4 release in March combined reasoning, multimodal, and coding capabilities that had previously been spread across distinct models.
Large 4 extends that consolidation into a much larger open-weight flagship. Its European training footprint and multilingual-data emphasis give Mistral an infrastructure and language-coverage narrative alongside its claim of leading US- and Europe-developed open models.
First-order effects
- The public preview gives developers a way to evaluate Mistral Large 4’s multimodal and agentic capabilities before Mistral’s planned October 27 weight release.
- Mistral can substantiate its model launch with a disclosed European deployment of 3,800 Nvidia Grace Blackwell GPUs and multilingual training data.
Second-order effects
- Mistral’s claimed open-model lead gives US- and Europe-based open-model developers a more specific benchmark comparison as buyers assess Large 4 ahead of the planned weights release.
- Nvidia gains a named European frontier-training reference for Grace Blackwell, while Mistral’s infrastructure disclosure raises the visibility of where large open models are trained.
Third-order effects
- If Mistral releases the weights as planned, open-weight competition will increasingly differentiate on the training and deployment stack—compute location, language coverage, and access path—not licensing alone.
- European model builders are beginning to frame local compute and multilingual training as product attributes, potentially making provenance part of enterprise model selection.
The trend: Open-weight frontier models are becoming competitions over controllable compute, multilingual capability, and staged access as much as model scale.