Mistral today released a preview of Mistral Large 4, a language model with 1 trillion (1T) parameters. The company reports the model operates with 49 billion active parameters and was trained on its own cluster of 3,800 NVIDIA Grace Blackwell GPUs.
Access and planned weight release
The preview is available via the Mistral API. Mistral says it will publish the model’s open weights “at the end of this month.”
Modes of operation and output behavior
Through the Mistral API the model supports two reasoning levels: “none” and “high.” In examples provided, the “high” setting produced a visually better output while surprisingly using fewer output tokens—2,717 tokens for “high” versus 3,275 tokens for “none.”
Performance and comparisons
On the "Artificial Analysis" (AA) benchmark, Mistral Large 4 scored 38 points, placing it just behind DeepSeek 4.1 Flash, a 552-billion-parameter model. This represents a substantial improvement over Mistral Large 3, which previously scored 9 on the same benchmark.
While Mistral Large 4 is not at the level of top-tier “Fable-class” models, its release suggests it is roughly six months behind the current frontier models.
Significance
The launch matters because of the model’s scale and the promise of publicly available weights, which can increase competition and access in the LLM ecosystem. The measured performance gain also indicates Mistral is making progress compared to its immediate predecessor.
