Mistral unveils Large 4, a trillion-parameter open-weight model

Mistral AI announced Mistral Large 4 on October 6, 2026, and opened a public preview through its API. The company describes it as a multimodal model with about one trillion parameters in total and 49 billion active parameters. Mistral says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European data centres. The preview is available now; the company says it plans to publish the model weights by the end of October. It has not yet detailed the final weight licence in this announcement.

The release is aimed at more than chat. Mistral highlights coding, tool-using agents, image understanding, scientific work, finance and legal tasks. Its post reports results including 82% on one Artificial Analysis cybersecurity test and 93% on Cybench. It also lists scores for agentic coding and work tasks. These figures are claims and evaluations cited by Mistral, not a substitute for independent testing. For some coding results, Mistral says the evaluator measured them privately before the benchmark harness was publicly released. Performance can also vary with task design, prompts and model access.

The timing of the weights matters. An API preview lets developers try the model without running it themselves, while published weights could let organizations inspect, adapt and host it on their own infrastructure, subject to the eventual licence and hardware requirements. Mistral says the preview runs on infrastructure in Europe and presents local control as important for regulated and security-sensitive work. That may appeal to companies seeking alternatives to relying only on closed, externally hosted systems.

Mistral also stresses the model’s cyber capabilities, including vulnerability analysis and security testing. Such tools can help defenders investigate software, but they can also be misused. The company says it is conducting additional red-teaming with security organizations and public authorities before releasing weights. That makes the release process and safeguards worth watching alongside capability scores.

Large 4 is a notable step in Mistral’s effort to compete at the top end while offering an open-weight route. The practical test will come after broader access: independent evaluations, the weight licence, deployment costs, and results on real workloads will show whether its scale translates into useful control and performance for developers and businesses.