Mistral AI opened a public preview of Mistral Large 4 on October 6. The model has 1 trillion parameters, of which 49 billion work on each token. Mistral says it will release the weights by the end of October, after testing the model with security experts and governments.
New to this? Read it in simple words
- Mistral AI from France has shown its biggest AI model, Mistral Large 4.
- Developers can try a preview now; free weights should follow by the end of October.
- Mistral says the model is very strong at cybersecurity work.
- An outside test puts it 64th of 225 models, above the middle.
- Parameters
- The numbers inside an AI model that it learns in training. More usually means a bigger model.
- Open weights
- A model whose learned numbers anyone can download and run on their own computers.
- Mixture of experts
- A model design where only some parts of the model work on each word.
What Mistral released
Mistral Large 4 has 1 trillion parameters, Mistral says. It is a mixture-of-experts model, so only 49 billion parameters work on each token. It reads text and images and writes text.
For now, developers can try it through Mistral’s own API. The list price is $1.36 per million input tokens and $4.18 per million output tokens.
Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centres. Pierre Stock, Mistral’s vice president of science, told TechCrunch that this was “two to three times less than our Chinese competitors”.
Artificial Analysis Intelligence Index: Mistral Large 4 Preview against the middle score of 225 models in its class.
Why the weights come later
Mistral says it will release the weights “by the end of the month”. Until then, cybersecurity firms, vetted partners and state authorities are testing a version with fewer limits.
“We’ll work with trusted partners and governments to make sure that the open source weights can be used to defend,” Stock told TechCrunch. An “upcoming release” page for the model is already on Hugging Face.
Mistral has not yet named the licence for the weights. It also promises more details on the design and training when they come out.
Claims and a first outside test
Mistral’s main claim is about cybersecurity. It says the model ranks among the top five on Artificial Analysis’s Cyber Index, and leads open models made outside China.
Artificial Analysis also tested the preview on its general Intelligence Index. The model scored 38, which puts it 64th of 225 models in its class. The middle score in that class is 26.
SiliconANGLE notes that the preview falls well behind leading closed models on popular coding tests. Artificial Analysis also found it very wordy: it used 200 million tokens on the test, against a typical 81 million.
Sources
Every fact in this story comes from the sources below. Open them to check our work.
- 1
- 2Primary source · October 6, 2026mistralai/Mistral-Large-4.0-1T05-A52B (upcoming release) Hugging Face
- 3
- 4Research · October 6, 2026Mistral launches open-source Mistral Large 4, details AI roadmap SiliconANGLE
- 5
- 6Research · October 7, 2026Mistral Large 4 Preview: Intelligence, Performance & Price Analysis Artificial Analysis
The size, the training setup and the cybersecurity ranking are Mistral’s own claims. The Intelligence Index score is Artificial Analysis’s own test of the preview, not of the final open model.
Mistral opened a preview of a 1-trillion-parameter model with open weights promised, though an outside test ranks it only 64th of 225. How we rate the mood