Mistral has announced Mistral Large 4, a new flagship model nicknamed “Le Chonk,” with full public model weights scheduled for October 27. The French AI company disclosed the model at a conference in Abu Dhabi on October 6 and is giving limited early access to cybersecurity specialists and government organizations before the wider release, according to Reuters and The Wall Street Journal.
The timing matters because Large 4 is not yet a normal public model release. Developers cannot currently download the final weights, inspect a definitive model card, or reproduce Mistral’s performance and safety claims. That distinction is important: today’s event is an announcement and controlled preview, while the meaningful open-weight milestone is expected later this month.
Independent reporting describes Large 4 as a multimodal mixture-of-experts model with roughly 1 trillion parameters in total and about 49 billion active for each token. Axios reports that Mistral trained it for around two months on 4,000 Nvidia Grace Blackwell GPUs in European data centers. The Wall Street Journal likewise reports a 4,000-GPU training run. Those specifications would make Large 4 materially larger in total parameter count than Mistral’s previous flagship generation while keeping only a fraction of the parameters active during inference, a design intended to increase model capacity without paying the full compute cost of a dense trillion-parameter model on every token.
Mistral is positioning the model as a high-end open-weight alternative to both closed U.S. systems and increasingly competitive Chinese open models. Reuters reports that the company says Large 4 performs especially well in cybersecurity and can handle work in coding, finance, geospatial analysis and chip design. Mistral also says its containment measures held during attempts to make the model escape its test environment. These are company claims relayed through press coverage, not independently reproduced benchmark or red-team results.
For organizations considering self-hosted or sovereign AI, the operational implication is potentially significant. A strong European open-weight model could give enterprises and public-sector teams more control over deployment location, model inspection and infrastructure choices. But “open-weight” does not mean cheap to run: a model with 49 billion active parameters will still demand substantial accelerator memory, high-bandwidth interconnects and careful inference engineering for serious production workloads.
The strongest reason to watch Large 4 is therefore not a benchmark headline but the control boundary it could change. If Mistral releases the promised weights on October 27 under terms that permit practical deployment, teams will be able to test the model on their own code, data and security workloads rather than relying on vendor demos. Until then, comparisons with other frontier open models remain provisional.
At publication check time, Mistral’s official news page did not yet contain a dedicated Large 4 article or model card. Reuters, The Wall Street Journal and Axios independently reported the announcement and broadly consistent release details. The model’s exact license, downloadable artifacts and reproducible benchmark package should be verified again when the public weights arrive.