Mistral Large 4 is the public preview model Mistral AI released through its official blog on October 6, 2026: one trillion total parameters with 49 billion active, natively multimodal, and the largest model the French company has built. The preview API went live on Mistral Studio the same day, while the open weights are due by the end of October — and before they drop, the model will be red-teamed by cybersecurity leaders, vetted partners and state authorities.
Where it claims strength: security, coding and vision
Mistral reports that Large 4 ranks among the top five models globally on the Artificial Analysis Cyber Index and leads open-weight models developed outside China by a wide margin. On a test that asks a model to reproduce a real open-source vulnerability and then patch it, it scores 82%, the highest of any model; it also solves 93% of the 40 challenges in Cybench. On coding, it posts 61.7% on DeepSWE v1.1 and a combined Coding Agent Index of 49.8%, ahead of DeepSeek V4 Pro and Qwen3.8 Max, plus 59.9% on AutomationBench, which measures business workflows. On visual grounding it edges GPT-6 Astra on Dense 200, 42% to 41%. Most of these numbers come from the vendor and its evaluation partners, so they should be read as launch claims until independent testing follows the weights release.
The stance on security work is deliberate. Mistral points out that several leading closed models score near zero on that vulnerability test because they refuse the task, while defenders often need exactly that proof-of-flaw work. Large 4 pairs strong cyber capability with self-hostable open weights: a reduced-moderation version is being red-teamed first, and after release organisations can run it on private cloud or on-premise under their own policies. Where a model like GLM-5.3 reaching Amazon Bedrock represents the cloud-distribution route, Large 4 bets on the weights-in-your-hands route, and the split between the two is widening.
European compute is the other half of the story
Large 4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the preview is served on the same infrastructure. Its training data spans more than 160 languages, including every official language of the European Union. Mistral frames the launch as the first milestone of the roadmap funded by its €3 billion Series D, says the model will anchor a new generation of specialised models, and notes that the reinforcement learning run behind the preview is still in flight.
For users, the practical read is twofold. API users can try it now, but this is a preview: the company says the model is still improving rapidly, with architecture and post-training details held back until the weights land. Teams planning to self-host must wait until late October, and a one-trillion-parameter model, even with 49 billion active, is a heavy deployment by any standard. The moment that matters is independent benchmarking after the weights drop — if third-party results land near the launch claims, the top tier of open-weight models gets a new name.
