Mistral Large 4 entered public preview on October 6, 2026, with API access through Mistral Studio. In its launch announcement, the company said it plans to release the model’s weights by the end of October. The preview and the planned downloadable release are separate milestones.
For teams evaluating the model today, the immediate route is the hosted preview. Mistral also says training is continuing, so results from this version should be read as measurements of a model still being refined.
On this page
Hosted access first, weights later
Mistral describes Large 4 as a natively multimodal model aimed at coding, tool-using workflows and document or image understanding. Its announcement presents demonstrations involving technical drawings, PDFs and geospatial images, alongside scientific and business tasks.
Those demonstrations show the uses Mistral is promoting. They do not establish how the preview will perform on a particular company’s documents, tools or error-handling requirements. A buyer comparing releases can follow availability changes in our AI model updates tracker; the distinction to watch here is between a hosted preview and weights that customers can deploy themselves.
Mistral says it will publish further architecture and post-training details with the weight release. The announcement’s planned self-deployment benefits therefore describe the coming release, rather than an option already supplied by this preview announcement.
Training and evaluation claims
The company says it trained Large 4 from scratch using 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers, and that the preview runs on the same infrastructure. This is Mistral’s account of the training and serving setup, rather than an independently audited capacity measurement.
Mistral reports results across software engineering, cybersecurity, scientific coding and professional document work. Some tests are internal; others use outside evaluation providers. The launch page also describes a blind coding assessment conducted with Surge AI, in which model identities were hidden from professional annotators.
The relevant comparison is always the named test and the version evaluated. A coding result, a document task and an image-grounding result measure different things. Readers can use our benchmark coverage to distinguish a published score from a general claim that a model leads the market.
What happens next
Before releasing the weights, Mistral says it is testing the model with cybersecurity specialists, vetted partners and state authorities. Those participants receive access with expanded cyber capabilities and reduced moderation, according to the company.
The next dated commitment is the end-of-October weight release. Mistral also says Large 4 will underpin specialized models, but this announcement gives no separate release date for those products. Today’s confirmed development is the public API preview; deployment choices and further technical disclosures depend on the subsequent releases.
Source: Mistral, “Introducing Mistral Large 4,” October 6, 2026. Checked October 7, 2026.
