Transformers 5.22.0 Adds Support for Mistral Large 4 and FP8 Weight Fine-Tuning

Hugging Face released Transformers 5.22.0 on October 29, adding support for Mistral Large 4 (Le Chonk) alongside the public release of the model weights. For teams using Transformers, this provides an official version to rely on for loading the model, handling image inputs, and subsequent fine-tuning. Earlier this month, Mistral offered an API public preview; this release makes the weights available for download and self-hosted deployment. Release notes, weight release announcement

Large 4 accepts text and images and outputs text, with support for a one-million-token context window. The model has about one trillion total parameters, with roughly 49 billion activated per token—or 52 billion when embeddings and the output layer are included. Its MoE selects four of 128 routed experts per layer, plus one shared expert. Sparse activation reduces computation per step, but engineering teams still need to budget memory for the full weights and runtime state; hardware requirements cannot be estimated from activated parameters alone. Release notes

Its attention combines MLA, which compresses KV representations, with a sparse position selection mechanism that selects up to 2,048 key positions per query. Hugging Face also provides deployment guides for FP8 and NVFP4; the NVFP4 path requires an NVIDIA Blackwell GPU. Note that the eight-H200 deployment recipe shown by the official team is configured for a context window of about 131K tokens. It does not directly validate serving a one-million-token context on the same hardware. Technical and deployment guide

Another substantial change in this release enables the use of FP8-quantized weights for adapter training, such as LoRA, by adding backpropagation for dense and grouped FP8 matrix multiplication. Under the stated 8K-token conditions, the release notes report that blockwise dequantization reduces peak memory for MoE-layer backpropagation from about 28.9 GiB to 2.8 GiB. This is a result for a specific operation, not an estimate of total training memory requirements. Quantization update

Teams adopting the model should also check the inference mode: the available options are low, medium, and high, with high as the default; setting it to none still uses high. Teams should measure quality on Chinese-language tasks, output tokens, and latency before choosing a serving configuration. The license permits modification and fine-tuning, but operating a model-as-a-service business requires a separate commercial license. Publicly available weights do not mean that all hosted uses are unrestricted. Inference settings, license details