ANALYSIS · COMPARISONS
Mistral Large 4: its own training figures put it above the EU AI Act's systemic-risk line
Mistral's 2 November 2026 technical report gives the duration, chip count and utilisation of Large 4's pretraining. Combined with Nvidia's specifications, that comes to roughly 1.3 × 10²⁵ operations, above the 10²⁵ line where the AI Act presumes systemic risk. Mistral's legal centre does not classify it so.

In brief
- According to the report, Large 4 was pretrained for 45 days on 3,840 Nvidia GB200 GPUs near Paris, about 4.15 million GPU-hours at 36% utilisation of peak BF16 throughput.
- Using Nvidia's dense BF16 figure of 2.5 petaflops per GPU, that is about 1.3 × 10²⁵ FLOP for pretraining alone, our calculation. The AI Act presumes systemic risk above 10²⁵.
- Mistral's legal centre lists Large 4 as a general-purpose AI model, with no systemic-risk label, and no public notification or rebuttal has been found.
Mistral's own numbers put Large 4 above the AI Act's compute line. Its legal page says nothing of the kind.
Forty-five days on 3,840 GPUs
On 2 November 2026, Mistral AI added a technical report to the Hugging Face repository of Large 4, its open-weight flagship. Pretraining took 45 days on 3,840 Nvidia GB200 GPUs "in a Mistral data center near Paris", about 4.15 million GPU-hours, with 36% model FLOPs utilisation measured against peak BF16 throughput, according to the report.
What a team deploying it should weigh
Nvidia lists 360 petaflops of BF16 for a 72-GPU NVL72 rack and notes that the dense figure is half the sparse one shown, so 2.5 petaflops per GPU, according to its specification page. Multiplied out, 4.15 million GPU-hours at 36% of that peak gives about 1.3 × 10²⁵ FLOP. That calculation is ours, and it covers pretraining only.
Ten to the 25th is the threshold above which a general-purpose model is presumed to carry systemic risk; its provider must notify the Commission "within two weeks", under Article 52(1), as the AI Office explains. On 16 November 2026, Mistral's legal centre still classed both the 6 October preview and the 29 October release as "general purpose ai model". No notification or rebuttal is public. The rules in force are set out in our AI Act timeline.
For a product team, the documents also disagree on hardware. The report runs the 606 GB NVFP4 checkpoint on four GB300 GPUs. The downstream documentation, version 2, asks for at least eight 80 GB H200s for FP8, and still says it "will be updated in case of open weights release".
If your staff use Vibe, inputs train Mistral's models by default unless they opt out; Enterprise accounts are opted out, per the help centre. How objection works across the EU is covered in our guide.
The gap in the record
Total compute is not published: no token count, no reinforcement-learning budget. Mistral may contest the presumption; only the Commission can settle it. More analyses sit in Lab, alongside OpenAI's Control Room under the AI Act.
Frequently asked questions
Is Mistral Large 4 a systemic-risk model under the EU AI Act?
Mistral does not classify it as one. But the training figures in its own report point to about 1.3 × 10²⁵ FLOP, above the 10²⁵ threshold that triggers the presumption. A provider can argue against the classification; the Commission decides.
What hardware do I need to run Mistral Large 4?
The technical report says the 606 GB NVFP4 checkpoint runs on four Nvidia GB300 GPUs. Mistral's AI Act documentation for downstream providers asks for at least eight H200 GPUs with 80 GB each for FP8 inference.
Does Mistral train on Vibe conversations?
By default, yes, unless you opt out, according to Mistral's help centre. Enterprise customers are opted out by default. A thumbs-up or thumbs-down with a comment lets Mistral use that exchange on any plan.
Share this article
Primary sources
- Mistral Large 4 (technical report) — Mistral AI (Hugging Face), published 2 November 2026 · accessed on
- Mistral Large 4: AI Governance, Legal Center — Mistral AI · accessed on
- Mistral Large 4 – Technical Documentation For Downstream Providers, v.2 — Mistral AI, published 29 October 2026 · accessed on
- Do you use my user data to train your Artificial Intelligence models? — Mistral AI (help centre) · accessed on
- NVIDIA GB200 NVL72 — Nvidia · accessed on
- General-Purpose AI Models in the AI Act – Questions & Answers — European Commission (AI Office) · accessed on
An article by the Rédaction Tests & Outils, published under the editorial direction of Bastien Hugon. Editorial method and use of AI



