Mistral has introduced Mistral Large 4, its largest model to date. The model features one trillion parameters and was trained inside the company's own European data centers.
Mistral released a public preview of the model, nicknamed le Chonk. The API is available through Mistral Studio. Model weights are scheduled to follow at the end of October.
Mistral Large 4 features one trillion parameters
According to Mistral, the model is natively multimodal with one trillion parameters, where 49 billion are active. On X, the company states it is the best open-weight model from the US or Europe across aggregated benchmarks. It claims state-of-the-art performance in cyber defense, manufacturing, and finance, alongside beating closed frontier models in visual grounding.
The Intelligence Index by Artificial Analysis gives the model 38 points. Its predecessor, Mistral Large 3, scored 9 points, and Mistral Medium 3.5 hit 14 points. Claude Opus 5.5 leads the index with 58 points.
Security tasks drive the core selling point
Mistral focuses heavily on IT security capabilities. In the Artificial Analysis Cyber Index, the model ranks among the top five models worldwide, according to the company. When asked to reproduce and patch a real vulnerability in open-source software, the model hits 82 percent. Mistral notes that Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task entirely.
Mistral positions this refusal gap as a selling point for enterprise users. The company states that defending software requires proving vulnerabilities are real. At the same time, the model blocks 93.3 percent of attacks in Lakera's B3 AI Security Benchmark.
Read nextReka AI Releases Rho-1 Omni-Model for Text, Images, Video, and RoboticsTraining infrastructure relies on European data centers
Mistral trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers. Training data covers more than 160 languages, including all official EU languages. The company plans a European variant operated entirely on its own under European law.
During the preview, Mistral charges $0.68 per million input tokens and $2.09 per million output tokens. Cached inputs cost $0.07. Documentation also lists prices at double those rates.
Mistral is red-teaming the model with security firms, vetted partners, and government agencies until the weights ship at the end of the month.
