On 6 October, French company Mistral introduced Large 4, also known as ML4. The model has one trillion parameters in total but uses 49 billion at a time. This is called a mixture of experts: only the parts suited to the task are activated.
What you can use today is a preview through Mistral Studio and the company's API – a service other software can call. It requires an account and paid usage. Mistral lists prices of $1.36 per million input tokens and $4.18 per million output tokens. The model can also work with images and is said to be trained for more than 160 languages, including all official EU languages.
The open part comes later. Mistral says it will release the model weights before the end of the month; Reuters gives 27 October as the date. More information about the architecture, evaluations and post-training is due at the same time.
Mistral describes Large 4 as particularly strong in areas including coding and cybersecurity. These comparisons mostly come from the company, so they should be read as launch claims. The practical distinction is simpler: developers can test the cloud version now, while local use must wait for the weights, licence terms and realistic hardware tests.