Mistral Large 4 Launches with 1 Trillion Parameters

Mistral has released Mistral Large 4, a one-trillion-parameter model designed primarily for IT security tasks. This model features 49 billion active parameters and scored 38 points in the Artificial Analysis Intelligence Index, significantly improving from its predecessor, Mistral Large 3, which scored just 9 points.

Mistral Large 4 aims to address vulnerabilities that competitors like Claude Opus 5.5 and GPT-6 avoid, with Mistral stating that its model excels in reproducing and patching software vulnerabilities. In benchmarks, ML4 achieved a 93.3% attack blocking rate in Lakera's B3 AI Security Benchmark and scored 82% in reproducing and patching vulnerabilities, outperforming Z.ai's GLM-5.2.

The model was trained using 3,800 Nvidia Grace Blackwell GPUs, with an additional 3,000 GPUs utilized for reinforcement learning post-training. During its training phase, ML4 generated 33 billion tokens daily. The model is set to be available on Mistral Studio, and its weights are expected to be released by the end of October 2026.

In a blind evaluation conducted by Surge AI, ML4 received a score of 3.74 out of 5 for code quality. Additionally, it scored 49.8% in the Artificial Analysis Coding Agent Index and 42% on the Dense 200 benchmark. Mistral's CEO, Arthur Mensch, emphasized that the model is designed to run in private clouds or on-premise, catering to organizations focused on security.

Mistral has secured an $830 million loan for a data center near Paris, aiming for a compute capacity of 200 megawatts by the end of 2027. The pricing for the model during the preview is set at $0.68 per million input tokens and $2.09 per million output tokens, with cached inputs costing $0.07. This follows Mistral's previous model releases, including Mistral Medium 3.5, which scored 14 points in the same index.