Sources disagree on the availability of weights for Mistral Large 4
The WIRED article describes Mistral Large 4 as available with open weights, while other sources describe their release as still planned. The public API preview has safety restrictions; selected partners are testing a version with reduced moderation.
The WIRED article calls into question the previous information that the release of weights for Mistral Large 4 is still to come: it describes it as an open-weight model that anyone can use and modify. By contrast, the TechCrunch article says the weights are not yet available and Mistral AI plans to release them in three weeks, after safety testing. The company announcement and the article in The Decoder expect the weights to be released at the end of the month. The WIRED article places the release of the final version at that time.
A public preview is available through the API on the Mistral Studio platform. According to the TechCrunch article, the public endpoint has safety restrictions. Mistral AI says selected security partners and government institutions receive a version with reduced moderation and enhanced cyber capabilities during testing. According to the company, the model is natively multimodal, has a trillion parameters and 49 billion active parameters, and focuses on areas including coding and cyber defense.
Sources also differ on evaluation results. The TechCrunch article says benchmark results are not yet available, while the article in The Decoder reports 38 points in the Artificial Analysis Intelligence Index compared with 58 points for Claude Opus 5.5 (Max). The company claims of leading performance in individual areas therefore cannot be equated with an overall lead.
According to the company, training took place on its own infrastructure in European data centers. The company announcement and the article in The Decoder report 3 800 NVIDIA Grace Blackwell GPUs, while the TechCrunch article cites a figure of 4 000 Nvidia GPUs. These figures describe the training infrastructure, not stated requirements for running the model yourself.
Why it matters
The availability of weights determines whether the model can only be used through the provider service or whether you can also plan your own deployment and modifications. For cyber defense, the safety restrictions of the specific access option also matter: the public API and the version tested by selected partners have different conditions.
Release card
Mistral Large 4
Mistral AI
- Specifications
- The model has 1 trillion parameters, of which 49 billion are active.
- Inputs
- According to the company, the model is natively multimodal. It supports analysis of documents, charts, satellite imagery and technical drawings, as well as code generation.
- Availability
- A public preview is available through the API on the Mistral Studio platform. According to the TechCrunch article, public access has safety restrictions; according to the company, selected partners and government institutions receive a version with reduced moderation and enhanced cyber capabilities during testing. The company announcement says the weights will be released at the end of the month, while the WIRED article already describes the model as available with open weights.
- Artificial Analysis Intelligence Index 38 bodů According to the article in The Decoder, the model scored 38 points in an index combining ten benchmarks, while Claude Opus 5.5 (Max) scored 58 points.
- Artificial Analysis Cyber Index – reprodukce a oprava zranitelnosti 82% According to the company, the model achieved a score of 82 % in a test of reproducing and fixing a real software vulnerability, whose results are also affected by competing models refusing tasks.
- Cybench 93% According to the company, the model solved 93 % of the tasks in a set of 40 exercises from cybersecurity competitions.
- Artificial Analysis Coding Agent Index 49.8% According to the company announcement, the combined score in a coding agent evaluation places the model ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.
- DeepSWE v1.1 61.7% The company announcement reports a score of 61.7 % in the DeepSWE v1.1 benchmark, taking the figure from the Artificial Analysis coding index.
- Coding, working with repositories and terminal tasks.
- Vulnerability research and fixes, and malware analysis.
- Analysis of documents, charts and technical drawings.
- The public API has safety restrictions; a version with reduced moderation is accessible to selected partners.
- The availability of weights for your own deployment cannot yet be reliably confirmed because the sources contradict one another.
Mistral AI positions the model as a competitor to closed models from US providers and open-weight models from China. It emphasizes customer control over operation and security tasks that, according to the company, some competing models refuse. The Decoder article also reports a lower score in the general Intelligence Index than for Claude Opus 5.5 (Max).
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
What was added since the original report
Verified updates
-
The public endpoint has safety restrictions.; The weights are scheduled for release in three weeks, after safety testing is completed.; According to the company, it trained the model on its own infrastructure with 4 000 Nvidia GPUs.; Benchmark results are not yet available.
- The public endpoint has safety restrictions.
- The weights are scheduled for release in three weeks, after safety testing is completed.
- According to the company, it trained the model on its own infrastructure with 4 000 Nvidia GPUs.
- Benchmark results are not yet available.
-
The model scored 38 points in the Artificial Analysis Intelligence Index.; The model was trained in the company’s own European data centers.; Selected security partners have access to a version with reduced moderation and enhanced cyber capabilities.; According to the article, the model reproduces and fixes software vulnerabilities.
- The model scored 38 points in the Artificial Analysis Intelligence Index.
- The model was trained in the company’s own European data centers.
- Selected security partners have access to a version with reduced moderation and enhanced cyber capabilities.
- According to the article, the model reproduces and fixes software vulnerabilities.
-
The new article says open weights are already available in the preview, while the event still describes their release as planned.; The article says the final version of the model is planned for the end of the month.; The model is optimized for coding and cyber defense.; The model is nicknamed Le Chonk.; Mistral claims it trained the model from scratch.
- The new article says open weights are already available in the preview, while the event still describes their release as planned.
- The article says the final version of the model is planned for the end of the month.
- The model is optimized for coding and cyber defense.
- The model is nicknamed Le Chonk.
- Mistral claims it trained the model from scratch.
Two audiences, two different impacts
What this means
For individuals
A programmer can try the preliminary version of the model for coding through the API on the Mistral Studio platform; the claimed capabilities are not yet consistently supported across all sources.
For a business
When planning tests, the security team must distinguish between the public API and access for selected partners with reduced moderation. Sources remain contradictory about the availability of weights for your own deployment.
Risks and complianceCheck the original
Event sources
confirmed by 3 independent sources · 4 publishers, 3 independent. We count feeds from the same owner only once.