Introducing Mistral Large 4

Mistral AI announced a public preview of Mistral Large 4, its largest and most capable model to date, on the Mistral Studio API. Weights are planned for release by the end of the month, pending red-teaming in real-world settings with cybersecurity leaders, vetted partners, and state authorities.

Mistral describes Mistral Large 4 as a 1 trillion-parameter natively multimodal model with 49 billion active parameters, with a hybrid instruct-and-reasoning MoE design and native fluency in more than 160 languages. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, where the preview is also served. Pricing listed is $1.36 per million input tokens and $4.18 per million output tokens.

On coding, Mistral reports scores of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4.0, a combined Coding Agent Index score of 49.8%, and a blind human evaluation with Surge AI where it ranked second of five models at 3.74, behind Claude Opus 5 (4.22) and ahead of Kimi K3, GLM-5.3, and GLM-5.2. On AutomationBench it scored 59.9%, and on AA-Briefcase it reached 1,393 Elo.

On cybersecurity, Mistral says Mistral Large 4 ranks in the top five globally on the Artificial Analysis Cyber Index, scored 82% on a vulnerability-reproduction-and-patching test, and solved 93% of Cybench challenges. Mistral states Claude Opus 5.5 and GPT-6 Astra score near zero on the same test because they refuse the task. On Lakera's B3 AI Security Benchmark it resisted 93.3% of attacks.

Mistral also reported visual grounding results of 42% versus 41% for GPT-6-Astra on Dense 200, state-of-the-art standing on SciCode-Verified among open-weight models, a KORA Benchmark score of 1.691, and outperformance of GPT-6-Astra on third-party legal and financial evaluations. It said the model is built on the same training, customization, and RL environment offered through Mistral Forge, and that work is funded by a €3 billion Series D round.