| Date | 7 Oct 2026 |
|---|---|
| Region | Europe / DACH |
Mistral on October 6, 2026 launched a public preview of Mistral Large 4, also called ML4 and “le Chonk,” and said the weights will be released by the end of the month. The post describes it as a 1 trillion-parameter natively multimodal model with 49 billion active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe. The preview API is on Mistral Studio today. The post does not name a licence, a context length, or a country of incorporation.
Who this affects
- For exporters: Not specified in the record.
- For investors: Stated preview pricing is $1.36 per million input tokens and $4.18 per million output tokens. Weights are not released with the preview.
- For counsel/compliance: The post says the model is “delivered through open weights” and pairs cyber performance with “open weights and self-deployment.” It does not name a licence.
- For policymakers: Training and the public preview are both placed in Mistral’s own datacenters in Europe. The post names cybersecurity leaders, vetted partners, and state authorities among the users it is written for.
The read
The release and the weights are not the same event. What is live on October 6 is a preview API. The open-weight claim is a date: “Weights drop end of this month” and “We will release the weights by the end of the month.” Until that drop, the control point is Mistral’s API on its own European infrastructure, not a file a customer can move. The scale figures on the page are 1 trillion total parameters, 49 billion active, and 3,800 NVIDIA Grace Blackwell GPUs. No licence name is attached to the promised weights.
What was the action?
Mistral launched a public preview of Mistral Large 4 and said the weights will follow by the end of the month.
What’s the penalty / scope, and how is it set?
No penalty. Scope in the post is a 1 trillion-parameter natively multimodal model with 49 billion active parameters, preview API pricing of $1.36 per million input tokens and $4.18 per million output tokens, and training on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European datacenters. The post says it is competitive with the strongest open-source models and, in visual grounding, surpasses frontier closed models. Those comparisons are the company’s own.
Who else is exposed?
NVIDIA is named as the GPU supplier class (Grace Blackwell). No customer is named. Sectors named: cybersecurity, finance, law, engineering, manufacturing, logistics, pharmaceuticals, science, shipping, and the public sector, plus cybersecurity leaders, vetted partners, and state authorities.
What happens next
The document says the weights will be released by the end of the month. The licence those weights will carry, and the context length, are not in this post.