Artificial Analysis lists Mistral Large 4 Preview at 38 on its Intelligence Index, one point below DeepSeek V4.1 Flash at 39, as Simon Willison reported. The result gives ML4 a comparison on one evaluation; it does not establish a general ranking across tasks or how the model will perform in a particular user’s workflow.
Willison also reports that Mistral’s API offers two reasoning levels for ML4: none and high. The available information does not establish how those settings affect quality, speed or cost.
The earlier version of this story said independent aggregate rankings had not established ML4’s comparative standing. That was accurate when the article was published at 18:32 UTC Tuesday; Willison published his report at 20:18 UTC Tuesday.
Mistral AI made Mistral Large 4, or ML4, available Tuesday through a public API preview, while downloadable model weights remain a future plan. Mistral says it expects to publish the weights later in October; Euronews reported Oct. 27 as the target date. Until then, customers can access the preview but cannot download the model to run themselves.
The launch gives Mistral a new entry in the open-weight race and supports its pitch for European control over AI deployment. Artificial Analysis’s index adds an independent comparison, but it does not validate Mistral’s company-reported benchmark scores or establish performance across tasks. Euronews reported on the launch and the planned release.
What has Mistral released so far?
The public preview is available through Mistral’s API in Mistral Studio. An API preview is the access Mistral has opened for ML4; it is not the same as making the model’s weights downloadable. A release could let customers download the model and run it themselves.
Mistral describes ML4 as natively multimodal, meaning it is designed to work with more than one type of input, and says it uses a mixture-of-experts architecture. The company identifies the model as having 1 trillion total parameters, with 49 billion active parameters. Its nickname is Le Chonk.
Mistral says the training data covered more than 160 languages, including every official language of the European Union. That is a company description of its training data, not an independent test of the model’s quality across those languages.
The company is promoting ML4 for coding, agent workflows, multimodal tasks and cybersecurity. It also says select partners and state authorities are testing the model in real-world settings, including with broader cyber capabilities than the public preview offers. Those tests do not mean the expanded capabilities are available to preview users.
Mistral says it trained the model from scratch over two months in European data centers. The company gives a total of 3,800 NVIDIA Grace Blackwell GPUs, while Euronews reports 4,000. The available reports do not reconcile the different counts.
How strong are ML4’s reported results?
Artificial Analysis’s Intelligence Index gives Mistral Large 4 Preview a score of 38, compared with 39 for DeepSeek V4.1 Flash. The one-point difference describes their results on that index; it does not show that either model performs better across every task or use case.
Mistral reports scores of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4. It also reports a 49.8% combined score on its Coding Agent Index. These are company-published results, not independently confirmed scores.
For cybersecurity, Mistral reports an 82% result on an Artificial Analysis Cyber Index test involving vulnerability reproduction and patching, and 93% on Cybench. The launch information does not establish independent verification of those scores.
The company says ML4 is competitive with leading open-weight systems. At launch, independent aggregate rankings had not confirmed its comparative standing. The new index result fills that specific gap, but it does not independently verify Mistral’s separate benchmark claims. Mistral’s test results can indicate how it performed under the reported conditions, but they do not by themselves establish its position across other evaluations or how it will work in a customer’s deployment.
Independent task-level results would help show how the index comparison relates to particular uses.
Free newsletter
Get the morning briefing
Start each day with the AI stories that matter and why — a short, free email from our newsroom.
What could downloadable weights change?
Making weights available could give customers the option to run and adapt ML4 themselves rather than rely only on the current API preview. That can offer more control over where and how a model is deployed, but the practical value depends on the terms and requirements of the eventual release.
As of Tuesday, the weights were not publicly downloadable. Customers therefore could not yet examine the files or determine the licensing terms and practical requirements for running ML4 on their own systems. Oct. 27 is a reported target, not a completed release.
A weights release would not, on its own, reveal all parts of model development. The availability of parameters does not establish that training data or methods will be public. The scope of disclosure and the license remain among the details to watch.
There is also a security trade-off. Public weights can be modified in attempts to remove safeguards; at the same time, Mistral argues that openness can help outside auditing and defensive cybersecurity work. Neither position, by itself, establishes how those competing effects will play out for ML4.
Does European hosting prove AI sovereignty?
Mistral presents European infrastructure and customer control over deployment as a sovereignty advantage. It says ML4 was trained in its European data centers and that the preview runs on European infrastructure. If the weights are published, customers could also have the option to operate the model on their own servers.
Those arrangements could give some organizations choices about where they host and deploy AI systems. They do not prove that Mistral is independent of non-European suppliers, technology or infrastructure. The launch materials do not map the full set of dependencies or spell out every requirement for self-hosting.
The distinction matters because access to a model and control over its full supply chain are different things. A European data-center location says where some training or hosting took place; planned downloadable weights could change where a customer runs a model. Neither fact answers every question about suppliers, financing or other dependencies.
Mistral’s launch also sits within competition between open-weight systems and more restricted offerings, including a contest involving US and Chinese developers. The new score is one comparative data point; task-specific evaluations and fuller information about the weights and supply chain would give a clearer picture of customers’ options.
The next concrete milestone is whether Mistral publishes the weights on its reported Oct. 27 schedule and what license and operating conditions accompany them. Further independent task-level evaluations would show how ML4 compares in specific uses. For now, the public has access to an API preview, not the downloadable weights.