Mistral Large 4 opens in preview at 1 trillion parameters

Mistral AI released Mistral Large 4 as its most capable model to date, but the launch splits availability into two stages. The model is live today only as a public preview on Mistral's cloud platform, while the weights that would let outside developers run it themselves are promised for later this month. That sequencing matters for open-source users, who cannot self-host the model until the second stage arrives.

Under the hood, Mistral Large 4 uses a mixture of experts design: it holds 1 trillion parameters but activates only 49 billion at a time, a technique meant to reduce the hardware needed per prompt. Mistral also described a training setup where its software stack ran tens of thousands of rollouts in parallel, producing 33 billion tokens per day, with just under half of those tokens feeding the training workflow. The company did not say how long training took, but it noted it never paused the run after creating the current model, expecting the same workflow to yield larger and more capable versions in the coming months, and eventually a series of models tuned for specific use cases.

On benchmarks, the picture is mixed. Mistral Large 4 earned a top-five score on the AA Cyber Index and 82% on patching open-source projects, ahead of open-source rivals, and it outperformed GPT-6 Astra by 1% on an image-object benchmark. It falls well behind frontier models like Astra on popular coding benchmarks, though it outperforms other leading open-source models on those and on knowledge-work tests. The open question is whether the promised weight release lands on schedule, and whether the follow-on models deliver the capability gains Mistral is projecting.