Downloadable weights follow the preview
Mistral Large 4 now has official FP8 and NVFP4 checkpoints, following its API preview. Hugging Face’s weight-release announcement is dated October 29. [1][2]
It accepts text and images and generates text. Mistral describes about one trillion total parameters, 49 billion active per token—or 52 billion including embeddings and output layers—and a one-million-token context. [2]
Check the full files and terms
The FP8 repository lists 54 weight shards totaling about 1.06 TB; NVFP4 lists 32, about 610 GB. These are stored file sizes, not a complete runtime-memory budget. NVFP4 targets Blackwell GPUs. [2][3][4]
The Mistral Ecosystem Open-Weight License requires a separate license before commercial use if the licensee or an affiliate operates a Model-as-a-Service business, with stated exceptions for internal use and Mistral’s official products. Read its definitions before offering customer access. [5]
Our take
For experienced technicians, the useful question is whether your workload justifies the hosting, and administration. Plan memory for weights, context and concurrent requests, then compare total operating cost with a hosted service. Keep your own evaluation set when checking vendor performance claims.
We inspected public documentation and file listings; we did not download weights, run the model, or benchmark it.