Mistral Large 4

Mistral Large 4, released October 6, 2026 as a public API preview, was the fourth generation of the flagship line Mistral opened in February 2024, and the lab's largest model at the time: a natively multimodal mixture-of-experts with 1 trillion total and 49 billion active parameters, a 512K-token context window and up to 256K tokens of output. Mistral pitched it at reasoning, coding and agentic work, said it had been trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the lab's own European datacenters, and promised open weights by the end of the month. Until then the preview was also red-teamed with security firms, vetted partners and state authorities, which got a less heavily moderated build with expanded cyber capabilities.

Cybersecurity was the lead claim. Mistral said the model was among the top five on the Artificial Analysis Cyber Index, reported 93% on Cybench and 82% on the index's reproduce-and-patch test, which it said was the highest of any model, and pointed out that several closed models scored near zero there because they refused the task. On coding it scored 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas Codebase QnA and 28.3% on Terminal-Bench 4.0 in Artificial Analysis runs made before the harness went public, for 49.8 on the Coding Agent Index, ahead of DeepSeek-V4-Pro-0813 and Qwen3.8-Max on Mistral's chart. It also posted 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase. Mistral framed the open weights and self-hosting as the point for security teams, who could run the model under their own policies without depending on a provider's refusals, and said the release would become the base for a new generation of specialised Mistral models.