Mistral Large 4: a large-scale model whose performance still needs to be measured

Announcement and access: a preview before the weights

Mistral announced Mistral Large 4 on October 6, 2026, and opened a public API preview in Mistral Studio that same day. The model is presented as the company’s largest system to date, with multimodal, coding and agentic capabilities. But that is not the same as making its weights available for download immediately: Mistral said it would release them at the end of October.

Initial access and weight availability are therefore separate stages. Anyone wanting to try the system now depends on Mistral’s API; self-deployment will only become an option once the weights are actually released and their terms permit that use. An announced release date does not, by itself, confirm that publication has happened or establish which licence will apply.

The company also says the model will be available in several regions, including a European option operated end to end by Mistral under European law. These are provider-announced features. Before treating them as operational guarantees, organisations should check the specific service terms, where processing takes place and the relevant contractual provisions.

One trillion parameters does not mean one trillion are active

Mistral describes Large 4 as a natively multimodal model with one trillion parameters in total and 52 billion active parameters. The active-parameter figure matters because not every parameter in a large-scale architecture necessarily participates in every operation. But those two numbers alone cannot establish actual cost, speed, memory requirements or response quality.

There is a discrepancy in the active-parameter figures: Mistral’s documentation lists 52 billion, while some supplied references give 49 billion. The available documentation does not explain whether the difference reflects a version, configuration or counting method. This article prioritises the figure in the primary announcement without treating the discrepancy as resolved.

Mistral says Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centres, and that the preview runs on the same infrastructure. The material available here includes no external audit verifying the number of GPUs, their location or the training configuration. This is therefore the company’s statement, not an independent verification.

The initial announcement also leaves technical details to be disclosed later: Mistral says it will provide more information about the architecture and post-training methodology alongside the weights. Until that documentation is available, it is best not to infer specific hardware requirements or assume that total model size translates directly into a practical advantage.

Benchmarks: promising results, but not universal comparisons

In results presented by Mistral, Large 4 scores 38 points on the Artificial Analysis Intelligence Index, according to the reading described in the available material. That can serve as a reference within that particular evaluation and version, but it is not a universal measure of intelligence. The result depends on which tasks are included, the conditions under which they are run and the timing of the evaluation.

Mistral’s announcement also reports results in coding, agentic workflows, finance, law, vision and science. Some figures come from the company’s own evaluations; others are attributed to external evaluators or benchmarks. A test having an independent name does not automatically make the way its result is presented independent: methodology, model version, settings and comparable data are all needed.

Model comparisons may cease to be equivalent when they use different versions, tools or execution budgets, or when safety filters prevent a model from completing a task. Mistral highlights favourable results in specific areas, but the material available here does not make it possible to independently reconstruct all the conditions behind those comparisons.

The cautious conclusion is limited: there are indications of substantial improvement over earlier Mistral models in some reported evaluations, but not enough evidence here to claim that Large 4 is generally superior to its rivals. A one-off ranking also does not necessarily predict performance on an organisation’s own documents, code, data and tools.

Agents and multimodality: announced capabilities that need real-world testing

Mistral presents Large 4 as capable of running general-purpose agents that gather information, use tools and complete multi-step workflows. The company also highlights its ability to analyse documents, charts and images, and to combine visual interpretation with tool-based actions. These are relevant goals for enterprise work, but the description does not justify assuming that the system can reliably complete every process without supervision.

In an agentic task, one mistake can cascade: misreading a document may lead to a faulty search and ultimately to an unintended action. Evaluating an agentic system therefore requires measuring more than whether it produces a final answer. Organisations should also assess the accuracy of each step, the quality of the sources it consults, how it handles failures and whether actions can be stopped or reviewed.

Multimodal capabilities also need to be tested on representative material. Analysing a technical drawing, extracting information from a PDF and finding objects in a large image are different tasks. A good result in a demo or benchmark does not guarantee accurate performance on an organisation’s particular formats, resolutions or documents.

For applications with significant consequences—for example, system changes, transactions or security decisions—organisations should limit permissions, log tool use and require human approval for irreversible actions. Automation can reduce workloads without giving the model full autonomy.

Cybersecurity: defensive potential and security limits

Cybersecurity is central to Mistral’s announcement. The company says Large 4 ranks among the top five models in the Artificial Analysis Cyber Index and reports results on exercises involving the reproduction and repair of vulnerabilities, Cybench and attack-resistance tests. The available material does not provide additional independent evaluations that would allow all these figures and comparisons to be confirmed here.

Mistral also reports a 93.3% result on Lakera’s public B3 AI Security Benchmark and describes internal tests involving malware analysis, vulnerability prioritisation and writing detection rules. Benchmark results should be distinguished from a guarantee that the model will be safe in production. A benchmark measures a defined set of cases, not every possible form of attack.

A model that can help find and reproduce vulnerabilities may also facilitate misuse. The ability to run released weights independently can give defensive teams more control, but it also leaves each organisation with greater responsibility for safeguards, logging, access and maintenance. Mistral says it is red-teaming the model with selected organisations before releasing the weights; that does not replace a public risk assessment or, by itself, define the final terms of use.

For a responsible deployment, the practical question is not only whether the model refuses malicious requests, but also how environments are separated, tools are controlled and outputs are reviewed. A useful evaluation should include authorised tests, human oversight and mechanisms that prevent model responses from directly changing sensitive systems.

What still needs confirmation before drawing conclusions

The final licence and the practical permissions it grants remain open questions in the information available here. Mistral announced that the weights would arrive at the end of October, but the initial announcement does not specify the licence. Without seeing it, organisations cannot assume they will be free to modify, redistribute or use the model in every way they might wish.

The complete deployment and cost conditions also remain to be confirmed. The supplied information includes conflicting API prices and does not establish one price that applies to every access method or region. Total costs may depend on usage, storage, infrastructure, support and the service terms in effect.

Announced availability is not the same as universal availability. The confirmed access route is the public preview; the weights were scheduled for release at the end of October, while regional options and European deployment are described as planned offerings. Their actual scope needs to be checked against up-to-date documentation.

In short, Large 4 merits attention because of its scale, its API offering and the announced plan to release the weights. But a robust assessment should wait for the licence, the promised technical documentation and more reproducible independent evaluations. Until then, performance and security figures should be read as reported results, not guarantees.

Keep exploring

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.