Mistral has released Mistral Large 4, which it describes as a 1 trillion-parameter natively multimodal model with 49 billion active parameters, built as a mixture of experts. It is in public preview through the Mistral Studio API from 6 October, and the company says it will release the weights at the end of this month.
The training detail is unusual in being stated at all. Mistral says the model was trained on 3,800 Nvidia Grace Blackwell GPUs in European datacentres, and that it spans more than 160 languages including every official language of the European Union. API pricing is $1.36 per million input tokens and $4.18 per million output tokens.
On benchmarks the company reports 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4 and 59.9% on AutomationBench, with 1,393 Elo on AA-Briefcase. Security is where it pushes hardest, claiming 93% of challenges in Cybench, 82% on a vulnerability reproduction and patching test, and resistance to 93.3% of attacks on the B3 Security Benchmark.
All of these are the vendor’s own figures on its own selection of benchmarks, and none has independent replication yet. The weight release at the end of the month is what would make them checkable by someone other than Mistral, which is the reason to wait before treating the security numbers as settled.
Source: Mistral AI, Introducing Mistral Large 4.