This is the English reading version of the Dutch news item Mistral Large 4 en NVIDIA mikken op CAD en simulatie.
Mistral has put Mistral Large 4 into public preview, nicknamed le Chonk. It is a multimodal model with roughly a trillion parameters, trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data centres. The weights become available at the end of October 2026.
The CAD connection sits in two places. Mistral itself mentions a human evaluation in which experts preferred the model in CAD and beta tasks over GLM-5.3. And NVIDIA and Mistral announced they are working together on engineering workflows.
Mistral’s own CAD claim
The announcement contains an internal evaluation in which annotators with knowledge of coding, CAD, finance, maths and physics compared Mistral Large 4 with GLM-5.3. The model was preferred in CAD and the beta tasks, and came close to GLM-5.3 on coding and finance. That is the maker’s own measurement, not an independent test, so take it as a signal and not as proof.
The collaboration with NVIDIA
Mistral and NVIDIA combine the model with CUDA-X and Warp for physics work, and with Omniverse for generating 3D scenes and SimReady assets. The goal is a closed loop between simulation, model development and design, so an agent not only draws a model but also starts the simulation and translates the result back into a change.
Alongside that, the two companies will publish open benchmarks and datasets for physics models, and build separate evaluations for agentic AI in CAE and EDA. Those benchmarks have to make the comparison between models fair, something that is measured differently per vendor today.
What you can expect from it
Mistral Large 4 is a general model, not a CAD program. The announcement is mostly about the training work and about tooling that still has to come. Anyone who wants to vibecad today has nothing to use it for yet; it gets interesting when the promised benchmarks exist and when it turns out that an agent can get from a design task to a simulated and improved model.
My verdict: the most interesting thing about Mistral Large 4 is not the collaboration with NVIDIA, but that the weights are opening up. A model you can run yourself means you are not tied to the API of a single provider, that you do not send your designs outside, and that you can keep working even if their prices change. Mistral says that itself as well: open weights exist to give organisations control over their own tools. The engineering workflows with NVIDIA are still promises; the open weights are a date.
Sources: the announcement of Mistral Large 4 and the collaboration with NVIDIA.