Mistral Large 4 is a public preview, not a downloadable open model yet

The model is available to try before the files are available

Mistral announced Large 4 on October 6 as a public preview. Developers can try it through Mistral Studio and the company’s API, but the same announcement says the model weights will arrive later in October. The current Mistral documentation still lists Large 4 as Public Preview and says the weights are coming soon. As of the October 24 check for this article, that is the clearest status: a hosted model is available; the downloadable checkpoint has not yet been announced as released.

The distinction matters because “open” is being used to describe the intended model release, not the immediate access path. A hosted preview lets developers send requests to a service that Mistral operates. Downloadable weights would let eligible users inspect the files and potentially run or adapt the model in their own environment, subject to the license and technical requirements Mistral publishes. Until those files appear, a user cannot treat the preview endpoint as self-hosting.

Large 4 is a very large multimodal model behind an API

Mistral’s current model card describes Large 4 as a general-purpose multimodal mixture-of-experts model with about 1.05 trillion total parameters and 52 billion active parameters. The October 6 press post rounds the total to one trillion. Both pages list a one-million-token context window, and the model card exposes features including function calling, document question answering, agents and conversations, structured output, and batching.

Those features make the preview more than a chat textbox. A developer can test how the hosted model responds to documents or calls tools through an application. But using function calling through an API is not the same as receiving a full self-hosted agent: the application still decides which tools are available, what permissions they have, and whether an action requires human approval.

Mistral says the public preview is served on infrastructure in its own European data centers, and its announcement says it was trained in the company’s European data centers. That is useful information about the provider’s stated hosting plan, but it does not answer every customer’s questions about retention, account terms, or the precise region used for each API request. Those details belong in the service documentation and contract for the account being evaluated.

Mistral says the preview is part of a security-testing phase

Mistral says it plans to release the weights after red-teaming the model in real-world settings with cybersecurity leaders, vetted partners, and state authorities. The company says those partners can use the same model with reduced moderation and expanded cyber capabilities. That is Mistral’s account of its testing process; the October announcement does not list every participant, test protocol, result, or release criterion.

For developers, the delay changes what can be evaluated now. The hosted preview can reveal whether the model’s responses fit a prototype and whether the API works with an existing integration. It cannot answer how a particular organization’s servers will run the weights, whether a quantized version will meet its latency target, or how much operational work self-hosting will require.

Mistral also says the reinforcement-learning run behind the preview is still in progress and that the model continues to improve. That makes version tracking especially important: an API result from early October may not match a later preview or the checkpoint eventually released. Teams should record which model ID and date they evaluated, then repeat their comparison after a material update.

The preview’s performance language remains mostly Mistral’s own

Mistral describes Large 4 as competitive with leading open models and says it performs especially well in areas such as cybersecurity, finance, and law. The company’s launch page includes preliminary evaluation charts, but its own announcement says the model was still improving and that a more complete set of results would come with the weights. The claim should therefore be read as the company’s assessment of a preview, not as a stable independent ranking.

Mistral’s current launch page reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. It also presents a blind coding evaluation and says the model continues to improve. These are Mistral’s reported results, not independent measurements; the company says additional benchmarks and methodology will accompany the weights. Compare scores only when model version, harness, and metric match.

For a team testing the API now, record the endpoint version and date, the prompt and tool setup, and the successful and failed cases. When the weights arrive, repeat a small set of tasks on the actual checkpoint. This separates improvements in the model from differences caused by a hosted endpoint, an agent framework, or a different evaluation script.

A weight release will answer new questions, not remove the old ones

If Mistral releases the weights, buyers will be able to inspect the license, file formats, supported runtimes, quantization options, memory requirements, and deployment instructions. Those documents will determine what “open-weight” permits for a specific use. The public preview page does not yet give a downloadable package or final self-hosting requirements, so it would be premature to assume the model can run on a particular cluster or be redistributed without conditions.

The model can still be useful before that release. Teams can test API behavior with synthetic or approved low-risk inputs, check whether its document and tool-calling features suit a workflow, and ask Mistral about data handling for the relevant plan. They should keep those findings separate from claims that require local weights. Preview access can show how a service behaves today; it cannot substitute for a local deployment test.

The API model card lists tools such as document Q&A and agent conversations, but a downloadable model is not the same bundle as the hosted service. The weights would provide the model itself; an organization would still need its own runtime, tool integrations, permissions, and safety review. A feature visible in Studio should not be assumed to appear automatically in a local deployment.

Budget the hosted API before expanding a pilot

Before expanding an API trial, estimate token costs and check the account’s limits. Mistral’s current Large 4 model card lists USD $1.36 per million input tokens, $0.14 per million cached input tokens, and $4.18 per million output tokens. The changelog records a launch offer of 50% off for two weeks; confirm any current promotion in the account instead of assuming that offer still applies. Mistral’s rate-limit guidance says limits vary by plan and model. Requests per second and tokens per minute are enforced independently for each organization; exceeding a limit returns 429, while batch jobs do not count against real-time limits. The admin usage and limits guide lists per-model request and token limits and allows workspace spending caps under the organization limit. Use a representative input/output mix and a defined workspace budget for the hosted test, then repeat the quality and resource checks on the released weights.