Mistral has launched a public preview of Mistral Large 4, its largest and most capable model yet. Nicknamed “le Chonk,” ML4 is a natively multimodal mixture-of-experts model with 1 trillion parameters and 49 billion active parameters. It combines instruction following, reasoning and agentic capabilities. The API is available now through Mistral Studio, with model weights due by the end of the month.
ML4 is aimed at software engineering, cybersecurity, finance, law, science and manufacturing. It scored 61.7% on DeepSWE v1.1 and 59.9% on AutomationBench. On one Artificial Analysis Cyber Index test covering vulnerability reproduction and patching, it reached 82%, while its Cybench score was 93%. Mistral also reports 42% on Dense 200 visual grounding, narrowly above GPT-6 Astra at 41%.
The model works across spreadsheets, documents, charts, technical drawings, PDFs and large geospatial images. Mistral says third-party evaluations place ML4 above GPT-6 Astra on legal and financial tasks, while Harvey’s Legal Agent benchmark shows it ahead of all open-source models. It also resisted 93.3% of attacks on Lakera’s B3 benchmark. Before releasing the weights, Mistral is red-teaming ML4 with cybersecurity leaders, vetted partners and state authorities using the same model with reduced moderation and expanded cyber capabilities.
The API costs $1.36 per million input tokens and $4.18 per million output tokens. ML4 will be offered in multiple regions, including a European deployment operated end to end by Mistral under European law. Private-cloud and on-premises use is planned for security teams, while training data covered more than 160 languages, including every official language of the European Union.
Mistral trained ML4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, which also serve the preview. The model uses the same training, customization and reinforcement-learning environment offered through Mistral Forge. At 3,000 GPUs, its current RL pipeline generates roughly 33 billion tokens per day, including about 16 billion trainable completion tokens after filtering and masking. The release is the first milestone backed by Mistral’s €3 billion Series D and will form the base for a new family of specialized models, pairing open weights with self-deployment and greater customer control.
Sources and related context #
- Mistral Large 4 model documentation : Supports: The official model documentation confirms public-preview status, multimodal capabilities and 49 billion active parameters.
- Mistral Forge enterprise model training : Related context: Mistral Forge explains the enterprise training, customization and reinforcement-learning platform referenced in the announcement.