Mistral Large 4 is Europe's trillion-parameter answer to US models that refuse security work
decoder ·
Mistral has unveiled a preview of Mistral Large 4, a one-trillion-parameter model trained in European data centers, positioning it as a sovereign alternative to US-based AI models. While the model scores 38 points on the Artificial Analysis Intelligence Index, significantly improving over previous versions, it still trails behind leading closed models like Claude Opus 5.5 by a wide margin. The primary competitive advantage highlighted is its capability in IT security, where it successfully reproduces and patches software vulnerabilities that competitors such as Claude and GPT-6 refuse to handle due to safety filters. Despite this edge in cyber defense, Mistral maintains high refusal rates for malicious prompts, aiming to balance utility with security. In coding tasks, the model ranks second among open-weight options, outperforming Deepseek and Qwen but falling short of Claude Opus 5. Additionally, Mistral emphasizes European sovereignty by training the model on independent infrastructure under EU law, supported by a record-breaking Series D funding round. Model weights are expected to be released at the end of October, accompanied by further architectural details.
decoder