Mistral launched a public preview of Mistral Large 4 on 6 October and says it is “significantly outperforming any open-weight model developed in the US or Europe”. The model has a trillion parameters, is natively multimodal and has 49 billion active parameters, and Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters. Only the hosted preview is available now, through the Mistral Studio API, and the weights follow by the end of the month.
The preview ranked second of five in a blind coding test, behind Claude Opus 5
Mistral reports that in a blind coding evaluation with Surge AI the preview ranked second of five models at 3.74 out of 5, behind Claude Opus 5 at 4.22. GLM-5.3 was the closest open-weight model at 3.60, followed by Kimi K3 at 3.59 and GLM-5.2 at 3.40. Its combined Coding Agent Index score of 49.8% puts it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.
The other figures come from Mistral’s own choice of benchmarks: 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4 and 93% of the challenges in Cybench. Mistral also says it is red-teaming a version with reduced moderation and expanded cyber capabilities with vetted partners and state authorities.
Running all trillion parameters would take more memory than the largest DGX Spark configuration
The 49 billion active parameters set how much arithmetic each token costs, but in a mixture-of-experts model any expert can be called on, so all of them have to be loaded or paged in. The size to plan around is the full trillion parameters.
For scale, NVIDIA’s DGX Spark page lists one 128 GB unit for models of up to 200 billion parameters, two units with 256 GB for up to 400 billion, and four units with 512 GB for up to 700 billion. The four-unit figure works out to roughly 0.73 GB of memory per billion parameters, which would put Large 4 at about 730 GB, above the 512 GB of the largest configuration NVIDIA lists. That is our arithmetic from NVIDIA’s table, not a requirement Mistral has published. The real figure depends on the quantization of the released weights and on context length, and Mistral has not stated either.
Until the weights ship, the only way to run the model is Mistral’s hosted API, priced at $1.36 per million input tokens and $4.18 per million output tokens. Mistral says it will publish architecture details and more benchmarks with the weights, and it has not yet stated the licence or the weight format.
Our view: at a trillion parameters the open weights will likely suit a datacenter or a multi-machine setup rather than a desktop, which makes the hosted preview the way most people will try Large 4 first.