NewsMistral2 min read
Mistral Large 4: Europe's Trillion-Parameter Open Model
Mistral unveils Large 4, an open-weight multimodal model trained in Europe: one trillion parameters, 160 languages and a very low API price.
Key points
- Mistral Large 4 has one trillion parameters, of which only 52 billion are active for each answer.
- It has been in preview in the Mistral Studio API since 6 October; the weights are due at the end of the month.
- It costs $1.36 per million input tokens and $4.18 per million output tokens.
- It was trained from scratch in Mistral's own data centres in Europe and covers more than 160 languages.
French company Mistral AI unveiled Mistral Large 4 on 6 October, its largest model so far. It is Europe’s most serious attempt to compete with OpenAI, Google and Anthropic, with one twist: it will be an open-weight model, so anyone will be able to download it and run it on their own servers.
Big, but efficient
Large 4 has one trillion parameters. But it uses a mixture-of-experts architecture: only 52 billion are active for each answer. That is how it can be huge and still relatively cheap to run. It is also natively multimodal: it understands text and images.
Price and availability
- API preview in Mistral Studio since 6 October.
- Open weights at the end of October, according to Mistral. The licence has not been announced yet.
- Price: $1.36 per million input tokens and $4.18 per million output tokens. For reference, GPT-6.1 Sol, OpenAI’s budget model, costs $2 and $10.
Made in Europe
Mistral trained it from scratch in its own data centres in Europe, on 3,800 NVIDIA GPUs, and will serve it from infrastructure it runs itself “under European law”. The training data spans more than 160 languages, including every official EU language.
In the results Mistral published, it stands out in cybersecurity (82% on the AA Cyber Index, the best score of any model according to the company) and in locating objects in images, where it slightly beats GPT-6 Astra. In coding (61.7% on DeepSWE) it trails the best closed models. These are the company’s own numbers: wait for independent evaluations.
What it means for you
- If your data must stay in Europe, Large 4 is probably the most capable option on European infrastructure.
- If you want to run a model on your own servers, wait for the weights at the end of the month. Note that a trillion parameters needs a lot of hardware, even quantised.
- If you pick API models on price, it is worth testing against OpenAI’s and Google’s budget models. Compare them in our model comparison.
