Mistral opens preview of Large 4 model

The public preview gives developers a way to test ML4 through an API on Mistral Studio.[1] The trained parameters are not described as already released: Mistral told the Wall Street Journal it planned to release them on October 27.[2] Mistral describes the model as natively multimodal, and says customers can use its open weights to gain more control over their AI.[1] The distinction matters for teams assessing a deployment: the current announcement offers preview access, while the planned parameter release is a later step.

Mistral’s published size figures are not consistent across its materials. Its blog describes ML4 as having 1 trillion total parameters and 49 billion active parameters; its documentation lists 1.05 trillion total, 52 billion active and a 1.6-billion-parameter vision encoder.[1][3] Mistral says the model solved 93% of Cybench’s 40 cybersecurity exercises, among the highest reported scores for an open-weight model.[1] It also reports scores of 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4.[1] CNBC reported that ML4 still lagged the frontier in areas including coding.[4]

Mistral says ML4 is particularly effective at cybersecurity, coding, manufacturing, finance and multimodal tasks.[4] The company says it worked with enterprises in finance, engineering, manufacturing, logistics, pharmaceuticals, science, shipping and the public sector to train the model.[1] Mistral says the model can be run on private cloud or on-premises infrastructure, and that it will be available in multiple regions, including an end-to-end European deployment operated under European law.[1]

Mistral says it could work with developers, cybersecurity experts and state authorities to evaluate ML4’s behavior.[2] It says these evaluation partners would receive the same model with reduced moderation and expanded cyber capabilities.[1] CNBC describes open models as modifiable and self-hostable, unlike leading closed systems from OpenAI and Anthropic.[4] That flexibility may make the planned weights relevant to teams seeking to adapt or host a model themselves, while the evaluation arrangement signals a separate process for assessing behavior in a less-moderated version.