What is new
Mistral Large 4, also called Le Chonk, takes text and images as input and generates text. Mistral describes it as multimodal, multilingual and built as a Mixture-of-Experts model.
Mistral lists about 1T total parameters and 49B active parameters, plus a context window of 1M tokens. The text decoder has 72 layers, a hidden dimension of 6,144 and 64 attention heads.
Benchmarks
Mistral reports 82% on the Artificial Analysis Cyber Index, a test that asks a model to reproduce a real vulnerability in open-source software and then patch it. Mistral says several closed frontier models score near zero on the same test because they refuse to attempt the work.
The model solves 93% of a 40-challenge suite drawn from security competitions on Cybench, and shows 93.3% attack resistance on Lakera's B3 benchmark. As of October 29, 2026, it is the fifth best open weights model on the Artificial Analysis Intelligence index, per the post.
Access, licence and hardware
Weights sit in the Mistral Large 4 collection on Hugging Face. GGUF versions come from Unsloth and the llama.cpp team.
The Mistral Ecosystem Open-Weight License, described as an MIT derivative, sets limitations for model-as-a-service offerings. A ready-to-deploy Inference Endpoints recipe runs on 8 x NVIDIA H200 GPUs, serves up to 131K tokens of context and handles up to 16 requests at a time. The NVFP4 checkpoint has a smaller memory footprint and needs NVIDIA Blackwell GPUs.
Mistral notes reasoning cannot be disabled, so reasoning_effort: none would still use high. The 2-bit quants run on a single Apple Silicon Ultra computer with 512 GB of unified RAM, while other setups need distributed inference. Prices are not listed in the post.
What it means for creators and sellers
Image input and a 1M-token context window may be useful for teams handling long briefs, product data or image-heavy material. Whether the output holds up on everyday marketing work is worth testing on your own files.
The licence limits matter if you plan to resell model access. If you only generate assets in-house, the open weights and llama.cpp support may make local trials easier to set up. Check current pricing separately, since this post lists none.
How this was made: the text was written by AI from the source material, checked against the source and reviewed by an editor. The illustration was generated by AI for AISTETIKA and is not a frame from the story.
