Mistral Large 4 open-weight release and day-0 support from Hugging Face, vLLM and SGLang
Article timeline
The articles in this story. Times are ET.
NVIDIA AIOfficialSGLang adds Day-0 support for Mistral Large 4 on NVIDIA Vera RubinAISGLang says it provides Day-0 support for Mistral Large 4, which it calls Le Chonk, and is the first to serve it on NVIDIA Vera Rubin. The quoted post describes it as a natively multimodal 1T MoE model with a 1M context window, and says Vera Rubin, NVFP4, and DSpark speculative decoding are supported on Day 0.
Hugging FaceOfficialHugging Face interviews Mistral on launch day of Mistral Large 4AIHugging Face spoke with Mistral's Pierre Stock and Sagar Vaze on the launch day of Mistral Large 4. The post says the model has open weights, was trained in Mistral's own data centers, and excels at cyber tasks. The full conversation covers cyber, training, evals and sandboxing.
vLLMOfficialvLLM adds day-0 support for Mistral Large 4 with FP8 and NVFP4 checkpointsAIvLLM announces day-0 support for Mistral Large 4, with Mistral leading the implementation alongside vLLM contributors and Red Hat, inferact and NVIDIA. The post links FP8 and NVFP4 checkpoints, with NVFP4 marked for Blackwell, and a technical deep dive on the blog. The image shows a disaggregated serving layout with prefill and decode stages, NIXL KV handoff, and cross-layer KV sharing and context parallelism.
- vLLMOfficialPickvLLM adds day-0 support for Mistral Large 4 with 1M context
AIvLLM says it offers day-0 support for Mistral Large 4, a trillion-parameter multimodal mixture-of-experts model with a 1M-token context window. The model targets coding, reasoning and cybersecurity tasks and ships FP8 and NVFP4 checkpoints with a built-in DSpark drafter.
- NVIDIA AIOfficialNVIDIA AI supports Mistral's new open Le Chonk model at launch
AINVIDIA AI congratulates Mistral on the next-generation Le Chonk, a 1T-parameter MoE multimodal model trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs. It offers day 0 support, including NVFP4 weights on Hugging Face and scaled distributed inference using Dynamo.
Lewis TunstallXPickHugging Face TRL adds day-0 support for post-training Le Chonk on one nodeAILewis Tunstall says TRL supports post-training Mistral's Le Chonk on a single node, using LoRA SFT on a 1T-parameter model with adapters on FP8 weights. He reports 32.7k tokens per step in about 8 seconds and 142 GiB per GPU on 8× B300, using expert parallelism and FSDP. The post links a training recipe on GitHub.
Julien ChaumondOfficialHugging Face releases Transformers v5.22.0 with Mistral Large 4 supportAIHugging Face releases Transformers v5.22.0, which includes support for Mistral's Large 4 model. The post's only other detail is the model's nickname, Le Chonk.
SGLangOfficialSGLang adds a cookbook guide for serving Mistral Large 4AISGLang says users can now serve Mistral Large 4 with SGLang, linking to a cookbook guide on its documentation site. The post gives no further details on performance, hardware, or setup steps.
- SGLangOfficialPickSGLang adds Day-0 support for Mistral Large 4 on NVIDIA Vera Rubin
AISGLang says it has Day-0 support for Mistral Large 4, also called Le Chonk, and is the first to serve it on NVIDIA Vera Rubin. The post says Vera Rubin, NVFP4, and DSpark speculative decoding are all supported on Day 0. Mistral describes the model as a natively multimodal 1T MoE with a 1M context window, and says it has open-weight state of the art for security work.
- Hugging Face BlogOfficialPickHugging Face introduces Mistral Large 4 with open weights and day-0 support
AIHugging Face says Mistral Large 4, nicknamed Le Chonk, is a multimodal Mixture-of-Experts model with about 1T total and 49B active parameters and a 1M-token context window. The model is released in FP8 and NVFP4 variants, with GGUF quants from Unsloth and the llama.cpp team, and day-0 support in transformers, vLLM, and llama.cpp. The post reports 82% on the Artificial Analysis Cyber Index and 61.7% on DeepSWE v1.1.
- vLLM BlogOfficialPickvLLM adds day-0 support for Mistral Large 4 with DSpark and NVFP4
AIvLLM announces day-0 support for Mistral Large 4, a 1-trillion-parameter mixture-of-experts multimodal model with 49 billion active parameters per token. The post reports about 2x output throughput on TP8 with the built-in DSpark drafter and about 1.5x more KV cache with the NVFP4 checkpoint, which stays within 2.1 points of FP8 on evaluated tasks.