The essentials
- On 6 October 2026, Mistral AI opened Mistral Large 4 in preview: roughly one trillion parameters, 52 billion active, with weights promised by the end of October.
- Trained in European data centres and offered with operated deployment in Europe: a real sovereignty argument, verified in architecture rather than in branding.
- “Open” does not mean “installable on any server”: at this scale, even compressed, you need data-centre-class infrastructure.
What Mistral announced
On 6 October 2026, Mistral AI opened preview access, including via its API, to Mistral Large 4. According to the vendor, it is a multimodal model of roughly one trillion parameters, with 52 billion activated on each forward pass, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centres in Europe.
The model weights—the file needed to run it yourself—are due to be released by the end of the month; specialist press cites 27 October. Mistral highlights top-tier performance among open models, a fully operated European deployment option under EU law, and the ability to run the model on premises for sensitive use cases such as cybersecurity.
The benchmark figures are those presented by the vendor itself. They warrant independent evaluation before being taken as settled.
Open, yes. Hostable on your premises, rarely.
That is the nuance headlines skip. A one-trillion-parameter model requires on the order of 1 TB of memory for weights alone at 8-bit precision and about 500 GB at 4 bits—before context cache and operational headroom. These orders of magnitude are a straightforward calculation from the announced parameter count: they place the model in the realm of multi-GPU data-centre servers, not an SME server room.
Two properties are too often conflated. “Open-weight” says whether the model can be downloaded, audited, and run under your own control. Total parameter count says whether it fits in your building.
The licence still needs reading
At the time of writing, the weights are not published and their terms of use are not either. According to published analyses, smaller models in the Mistral family (“Small”, Devstral, Magistral Small) are released under Apache 2.0, which allows commercial use: verify model by model before any project.
Before choosing a model to run in-house, five checks:
- Use case first: document search, summarisation, extraction, assistance. Each task has its own quality threshold.
- Memory actually required: compressed weights, context cache, and operational headroom, before talking about hardware.
- The exact model licence: commercial use, redistribution, any restrictions.
- Evaluation on your documents: an internal test set beats a public leaderboard.
- Reversibility: ability to change models without rebuilding the entire surrounding stack.
Sources
This note was drafted with AI tools from the cited sources, then reviewed and published under the responsibility of Jordan FOUASSIER, publication director. We summarise the facts and add our reading; source text and images are not reproduced. An error? Write to contact@aigyrosgroup.com.