Mistral AI has opened a public preview of Mistral Large 4 (ML4), a 1-trillion-parameter, natively multimodal, open-weight model with 49 billion active parameters. The preview API is available now through Mistral Studio, served on Mistral's own European datacenter infrastructure; full model weights are scheduled for release by the end of the month.
The company reports ML4 at 59.9% on AutomationBench, a 657-task benchmark covering agentic use of Gmail, Google Sheets, Slack and Salesforce, ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro on that specific measure. On AA-Briefcase, a benchmark for long-horizon knowledge work such as spreadsheets, slides and PDFs, it reaches 1,393 Elo, also ahead of DeepSeek V4 Pro.
Mistral also reports strong results on agentic coding (61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA), cybersecurity (93% on Cybench, 82% on a vulnerability reproduction-and-patch test), and legal and finance tasks evaluated by third-party evaluator vals.ai, where the company says ML4 exceeds GPT-6-Astra on both. On security robustness, Mistral reports ML4 resists 93.3% of attacks on Lakera's public B3 benchmark and tops its own robustness-to-prompt-injection evaluations among open-weight models tested.
Pricing for the preview API is listed at $1.36 per million input tokens and $4.18 per million output tokens. Full weights, additional architecture details, benchmarks and post-training methodology are promised alongside the end-of-month release. Mistral frames ML4 as the first product milestone funded by its €3 billion Series D round and says the model will serve as the base for a further generation of specialized models.