Mistral AI makes a public preview of Mistral Large 4 available via API
The public preview of Mistral Large 4 is available via API on the Mistral Studio platform. According to Mistral AI, the model has a trillion parameters, of which 49 billion are active, and is natively multimodal. The company plans to release the weights at the end of the month.
Mistral AI has made a public preview of Mistral Large 4 available via API on the Mistral Studio platform. It plans to release the weights at the end of the month. Until then, according to the announcement, it is continuing security testing with vetted partners and government agencies; they are to receive access with reduced moderation and enhanced cybersecurity capabilities.
According to the company, the model has a trillion parameters, of which 49 billion are active, and is natively multimodal. The company says it trained the model from scratch on 3 800 NVIDIA Grace Blackwell GPUs in its own European data centers, where the public preview also runs. According to the announcement, the training data covered more than 160 languages, including all official languages of the European Union.
The company reports capabilities in coding, agent workflows and multimodal understanding. In the announcement, it reports an 82% success rate in a test involving the reproduction and repair of a real software vulnerability within the Artificial Analysis Cyber Index, and the completion of 93 % of the 40 tasks in the Cybench suite. For coding, it reports a score of 61.7 % in the DeepSWE v1.1 test. Claims of superiority over other models come from the company announcement. Details can be found in the source article.
Why it matters
API access allows developers and security teams to evaluate the model on their own tasks during the public preview. According to the company, the planned release of the weights would also allow deployment in a private cloud or on their own infrastructure, which is important for organizations that need to manage access and security policies themselves.
Release card
Mistral Large 4
Mistral AI
- Specifications
- According to the company, the model has a trillion parameters, of which 49 billion are active.
- Inputs
- According to the company, the model is natively multimodal and supports visual understanding.
- Availability
- The public preview is available via API on the Mistral Studio platform. The company plans to release the weights at the end of the month.
- Artificial Analysis Cyber Index – reprodukce a oprava zranitelnosti 82 % The company reports an 82% success rate in a test involving the reproduction of a real vulnerability in open-source software and its subsequent repair.
- Cybench 93 % According to the announcement, the model solved 93 % of the tasks in a set of 40 exercises from cybersecurity competitions.
- DeepSWE v1.1 61,7 % The announcement reports a score of 61.7 % in a software engineering benchmark and refers to results reported by Artificial Analysis.
- SWE-Atlas-QnA 59,4 % The announcement reports a score of 59.4 % in a benchmark focused on understanding software repositories.
- Terminal-Bench 4 28,3 % The announcement reports a score of 28.3 % in a benchmark for working in the terminal.
- Coding and working with software repositories.
- Agents that gather information, use tools and prepare outputs.
- Malware analysis, vulnerability prioritization and the creation of detection rules, according to internal company tests.
The company claims the model is competitive with the strongest open models and outperforms models with open weights developed in the US or Europe. In the combined Coding Agent Index, it reports a higher score than those for DeepSeek V4 Pro 0813 and Qwen3.8 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 new article states that 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 states that 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 developer can use the API to try the model for assistance with coding and working with repositories, and assess the quality of the results for their specific tasks.
For a business
Security teams can begin evaluating the model for vulnerability analysis and the creation of detection rules. However, when deciding whether to run it on their own infrastructure, it is important to note that the release of the weights is still only planned.
Risks and complianceCheck the original
Event sources
clearly official source · 1 publisher, 0 independent. We count feeds from the same owner only once.