Breaking02 / IT
Mistral publishes Large 4 weights under modified MIT licence
Commercial use is allowed, but a model-as-a-service business needs a separate commercial licence.

Mistral released the weights of Large 4 on 29 Oct 2026 under the Mistral Ecosystem Open-Weight License, and Hugging Face said a 2-bit quantised version runs on a single machine with 512GB of unified RAM.
Hugging Face lists about 1T total parameters, with 49B active per token. The model takes text and images as input and has a context window of 1M tokens. It comes in two official variants, FP8 and NVFP4.
Hardware. In Hugging Face's test, the Q2_K quantisation generated 15 tokens per second on a single M3 Ultra. The 2-bit versions run on a single Apple Silicon Ultra machine with 512GB of unified RAM. The others need distributed inference, which splits a model across several processors or machines.
Mistral said it carried out further reinforcement learning after the public preview to improve the model. Large 4 has three reasoning modes, with low replacing the None mode. According to Hugging Face, setting reasoning_effort to none still runs the model at high.
Licence. The release notes for transformers v5.22.0, a Python library for running and training models, describe the licence as a modified version of MIT. It covers both commercial and non-commercial use. A model-as-a-service business, which sells access to a hosted model, requires a "separate commercial license" with Mistral.
Hugging Face said TRL, a fine-tuning tool, supports Large 4 from release day. The transformers release notes state that adapter training such as LoRA works on FP8 weights from v5.22.0.
According to the release notes, training took place in Mistral's own data centres in Europe on 3,800 NVIDIA Grace Blackwell GPUs. The notes do not give redistribution terms or the sources of the training data.