A trillion parameters to download: Mistral releases the weights of its top model Large 4
The French AI company has published the weights of its largest model. Companies can run it on their own servers – but providers of AI services need a separate licence.
Photo: Slush / Wikimedia Commons, CC BY 4.0 (cropped and resized)
Video: YouTube · Mistral · youtube.com
French AI company Mistral AI on Thursday published the weights of its language model Mistral Large 4 on the Hugging Face platform. The model has around one trillion parameters, of which only a portion – about 52 billion, according to Mistral – is active per task. It understands text and images and outputs text and code. Mistral had initially made Large 4 available only through its API in early October; in its launch video the company presents the model, nicknamed "Le Chonk".
The files total around 1.06 terabytes; according to Mistral, a quantised version runs on four Nvidia B200 or B300 GPUs. The focus is on software development and cybersecurity. On Terminal-Bench 4.0, Large 4 scores 41 percent according to the company, putting it just ahead of the Chinese model GLM-5.3. Smaller distilled versions – including one for laptops – are to follow in the coming months.
The licence is a modified MIT licence with one important restriction: anyone offering the model as a service to third parties needs a separate commercial licence from Mistral. Explaining the release to The Next Web, VP of Science Pierre Stock argued that AI will become like electricity and that no company should depend on a provider that can cut off its access.