Mistral launched a public preview of Mistral Large 4. It is a 1 trillion-parameter natively multimodal model. The model is available via API on Mistral Studio today. Weights will drop by the end of October 2026. It was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European datacenters. The model excels in coding, agentic workflows, and multimodal understanding. It leads open-weight models in cybersecurity, finance, and law benchmarks. It supports over 160 languages, including all EU official languages. It will be deployable in multiple regions, including a European sovereign option. Red-teaming is ongoing with cybersecurity leaders and state authorities. The model resists 93.3% of indirect prompt injection attacks on B3. It refuses malicious cyber requests at a higher rate than other open models. It scores 61.7% on DeepSWE v1.1 and 59.9% on AutomationBench.
What changed
Mistral Large 3 was the prior flagship model with no public preview API for the next version.
What it unlocks
Running advanced cybersecurity, coding, and multimodal tasks with open weights and self-deployment.
- 1 trillion parameters
- 49 billion active parameters
- 3,800 NVIDIA Grace Blackwell GPUs
- 160+ languages in training data
What you need to act on it
- public preview API access
- weights release by end of October 2026
Sources