Mistral Large 4 scores 82% on a cyber test and gets open weights

Mistral opened a public preview of Large 4, its largest model, through its API. The model has 1 trillion parameters, and 52 billion are active. Mistral will release the open weights, the files that let a company operate the model, by the end of the month. The 82% score on a cyber test comes from an independent index.

What is new

Large 4 works with text and images. Mistral trained it on 3,800 NVIDIA GPUs in its datacenters in Europe. Until the weights come out, vetted partners and state authorities test a version with reduced moderation and more cyber skills.

The numbers

• Cyber Index of Artificial Analysis: top five of all models, and 82% on the find and patch test.
• Cybench, 40 security exercises: 93% solved.
• DeepSWE coding test: 61.7%.
• Blind human rating of coding: 3.74 of 5, second of five models.

Mistral says that Claude Opus 5.5 and GPT 6 Astra score near zero on the cyber test, because they refuse the task.

What is not known

Mistral gives no license terms, no model design and no more benchmarks at this time. It says it will give them before the weights come out.

Insight: Refusals by closed models make a market for open weights

Mistral says that Claude Opus 5.5 and GPT 6 Astra score near zero on its cyber test, because they refuse the task. An earlier report gives the Large 4 preview price as $1.36 for each million input tokens, against $5.00 for GPT-5.5. This reads as a move to sell open weights as the answer to the safety filters of closed models, at a lower price. Read the license terms first.