Mistral Large 4 enters public preview with a 1-million-token context; open weights promised this month
Oct 6, 2026
Paris-based Mistral AI opened a public preview of Mistral Large 4 on October 6, 2026. It is the company’s new flagship model, and Mistral says the full model weights will be released for download by the end of October.
Key facts
- Architecture: mixture of experts (MoE). Only a part of the model works on each word, which keeps it faster and cheaper to run than its full size suggests.
- Size: reports put it at roughly one trillion parameters in total, with around 50 billion active per token. Mistral has not confirmed every detail yet.
- Context window: up to one million tokens – enough to read several long books or a big code base in one go.
- Multimodal: it understands images as well as text.
- Where to try it: as a hosted preview in Mistral Studio, with function calling, structured output and document question answering.
Why “open weights” matter
With open weights, companies and developers can download the model and run it on their own servers instead of sending data to someone else’s cloud. That is attractive for banks, hospitals and governments with strict privacy rules. Mistral says the delay before the weight release is for extra safety testing and red-teaming with cybersecurity and public-sector partners. The license terms have not been published yet.
Can I run it on my PC?
Not the full model: a trillion-parameter model needs data-center hardware. Home users can try the hosted preview, or keep using smaller open models that run on a normal graphics card. Independent benchmark results are still missing, so treat early performance claims with care until they are tested.
Sources
- Let’s Data Science: Mistral opens public preview of Large 4
- Tech Startups: Tech news weekly roundup (October 9, 2026)
Tags: Mistral Large 4, open weight AI model, open source AI, 1 million token context, mixture of experts, Mistral AI, local AI