|
French superintelligence company Mistral has launched a public preview of its latest Mistral Large 4 (ML4) model, a 1-trillion-parameter natively multimodal system, with the company set to release its weights by the end of the month as it scales up its European computing infrastructure.
Mistral said ML4 has 49 billion active parameters and is its largest and most capable model to date. The company said the model demonstrates exceptional performance across coding, agentic workflows, and multimodal understanding and critical enterprise applications including cybersecurity, finance and law. The company said the model has already shown performance competitive with leading open-source models and claimed it significantly outperforms open-weight models developed in the US and Europe. Mistral also said ML4 is state-of-the-art among open models on several enterprise workloads, while exceeding some closed models in areas such as visual grounding. The public preview is currently available through Mistral Studio, while the model weights are scheduled to be released by the end of October. Until then, Mistral said it is testing the model in real-world settings with cybersecurity leaders, vetted partners and state authorities. The company said ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters. It said the model will be available across multiple regions, including a European deployment operated independently by Mistral and under European law. Mistral is also positioning the model around greater control for organisations, particularly in cybersecurity. The company said ML4 combines advanced cyber capabilities with open weights and self-deployment, allowing organisations to run security work under their own policies. The company reported an 82 per cent score on one cybersecurity test involving reproducing and patching a real software vulnerability, while scoring 93 per cent on Cybench. In coding, it scored 61.7 per cent on DeepSWE v1.1 and 28.3 per cent on Terminal-Bench 4. Mistral further said ML4 is the first milestone under its EUR 3 billion (USD 3.49 billion) Series D funding and that expanded European computing capacity will support further training. The model is showing no signs of saturation, the company said, adding that it expects large and rapid improvements in the weeks and months to come. Mistral said ML4 will also serve as the foundation for a new generation of specialised models aimed at industries and workloads covered by its customers. (ANI)
|
