Observed facts: Mistral describes Large 4 as a natively multimodal 1-trillion-parameter model with 49 billion active parameters. The company says the preview API is available in Mistral Studio and that weights are due by the end of October. Mistral also says it trained the model on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers.
Mechanism: A hosted preview permits early use, but downloadable weights would allow organizations to run the model under their own infrastructure and policies. That matters especially in security work, where Mistral argues provider-level refusals can obstruct legitimate vulnerability research and incident response. The claimed capability therefore matters operationally only if deployment access, model behavior, and safeguards hold up in practice.
Why now: The preview is available now, while the promised weights are still pending. This creates a near-term gap between a vendor-operated API and the autonomy Mistral associates with self-deployment.
Implications: The first implication is that customers should treat the weights release and independent evaluation as the key milestones for judging whether the sovereignty proposition is operationally real. Mistral reports strong cyber, coding, and agentic results, but these are vendor claims; the post notes some coding evaluations were private before public harness launch.
Evidence boundary: The source is Mistral’s announcement, so its performance comparisons and deployment claims are not independent verification. It establishes what the company announced, not that customers have already received weights or reproduced the results.
Watch condition: The first watch condition is whether Mistral releases weights by the stated month-end deadline and whether independent evaluations reproduce the cited performance on public benchmarks.