Mistral AI
Open and portable generative AI for devs and businesses
5.0•41 reviews•4.1K followers
Open and portable generative AI for devs and businesses
5.0•41 reviews•4.1K followers
- We’re committed to empower the AI community with open technology. Our open models sets the bar for efficiency, and are available for free, with fully permissive license.
- Our optimized commercial models are designed for performance and are available via our flexible deployment options.
This is the 24th launch from Mistral AI. View more
Mistral Large 4
Launching today
Mistral Large 4 (aka Le Chonk) is Mistral's largest model yet: a natively multimodal mixture of experts model with 1T total and 49B active parameters and a 1M token context window. Mistral says it is the strongest open weight model from the US or Europe, and it ranks top 5 on the Artificial Analysis Cyber Index. It was trained in Mistral's own European datacenters. The preview API is live today on Mistral Studio, with open weights due by the end of October.
Free Options
Launch Team
ChatGPT by OpenAI
Claude by Anthropic
Perplexity for Mac
Hey PH community 👋
Excited to hunt Mistral Large 4 today!
Mistral just shipped its biggest model ever, and the team nicknamed it Le Chonk. It has 1 trillion parameters with 49 billion active, it understands text and images natively, and Mistral says it's the strongest open weight model built in the US or Europe. Mistral AI posted a short launch video on X here.
Here's some thoughts from the community:
https://x.com/simonw/status/2107479817780400577
https://x.com/ArtificialAnlys/st...
The cyber results are what stood out for me. Mistral Large 4 ranks in the top five on the Artificial Analysis Cyber Index, and on the test where a model has to reproduce a real vulnerability and then patch it, Mistral reports a score of 82%, the highest of any model. Several closed models score close to zero on that test because they refuse the task.
Heavy visual work is the other one I'd test first. It handles engineering drawings, dense charts, and gigapixel satellite images, which is where a lot of models still fall apart.
For builders, the weights land at the end of October so you can run it on your own servers. Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and the full benchmarks are in the announcement here.
Biggest win? This latest Mistral launch got the Pieter Levels vote of approval ;)
You can try the preview API on Mistral Studio right now: https://console.mistral.ai/
Give it a spin and let the team know what you build with Mistral Large 4 🚀
PS: Tagging @arthurmensch and @sophia_yang_mistral if you're here for any additional comments/ thoughts.