Mistral Large 4 is available as a public preview, while the release that would let organisations operate its weights themselves remains a plan for later this month. That distinction is central to Mistral’s announcement: access through Mistral Studio exists now, but the promised open-weight deployment options should not be described as already delivered.
AI News reports that the natively multimodal model has one trillion parameters, with 49 billion active parameters. Mistral’s account combines training details, performance claims and cybersecurity testing. Its reinforcement-learning work on the preview is still continuing, making the announcement a description of a model under development as well as a set of proposed deployment choices.
Preview access comes before control of the weights
Mistral says it trained Large 4 from scratch using 3,800 NVIDIA Grace Blackwell GPUs in its own European datacentres. The preview operates on that infrastructure too. The company reports training data covering more than 160 languages, including all official EU languages. These details describe Mistral’s stated development and hosting arrangements; they do not independently establish performance across every language represented in the data.
The weights are scheduled for release by the end of October 2026, together with architectural information, further benchmarks and the post-training methodology. Mistral intends the open-weight release to support private-cloud and on-premise operation, where organisations could apply their own policies. The preview API and those intended deployment arrangements are different stages of access, not interchangeable descriptions of what users can obtain today.
Cybersecurity testing also involves differentiated access. Mistral says security leaders, vetted partners and state authorities will test the same model with reduced moderation and expanded cyber capabilities. The company argues that provider-level refusals can impede legitimate security research and incident response. That is its rationale for the approach, not a finding that expanded capabilities are harmless in every setting.
Security scores describe tests, not universal protection
According to Mistral, Large 4 sits among the top five models on the Artificial Analysis Cyber Index. The company reports an 82% result on a task requiring reproduction of a real vulnerability in open-source software followed by a patch, which it describes as the highest model score on that test. It also reports solving 93% of Cybench’s 40 security-competition exercises.
These claims concern defined evaluation tasks. They should not be turned into a general probability of successfully defending a business. Mistral’s internal testing found uses in malware analysis, vulnerability prioritisation and writing detection rules, but the source does not establish that those uses consistently deliver the reported benchmark performance in operational environments.
The same caution applies to the company’s reported resistance to 93.3% of attacks on Lakera’s public B3 AI Security Benchmark. That figure represents performance within the benchmark, not evidence that almost every prompt-injection attack will fail. The distinction is particularly important alongside the planned open-weight release: reported test results inform assessment, while deployment choices and an organisation’s policies remain separate considerations.
Coding evaluations sit alongside unfinished training
Mistral cites software-engineering results of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, with a combined Coding Agent Index score of 49.8%. A separate blind evaluation with Surge AI used professional annotators rating coding output from one to five. Large 4 Preview came second among five models at 3.74, behind Claude Opus 5 at 4.22. The different evaluation formats should not be collapsed into a single claim of overall superiority.
Business workflows provide another test: Mistral reports 59.9% on AutomationBench, covering 657 workflows across applications including Gmail, Google Sheets, Slack and Salesforce. It also demonstrated visual tasks involving drawings, PDFs and satellite imagery. These are additional parts of the company’s performance account, rather than independently verified promises for every workflow or document.
Behind the preview, Mistral describes a reinforcement-learning environment combining scientific problems, conversation, factuality, safety alignment and extended tool use. Verification includes reward models, unit tests, model judges and static checks. At a reported scale of 3,000 GPUs, training produces about 33 billion tokens daily, including roughly 16 billion trainable completion tokens after filtering and masking. With that run continuing, both further technical disclosure and the planned weight release remain important next steps in evaluating the offering.
