Mistral releases Large 4 open weights despite risks
๐ก๏ธ Mistral AI has published the open weights for its trillion-parameter model Large 4 (Le Chonk), weeks after a public preview where the company monitored cybersecurity-related usage. The firm says the model refused most live attack prompts but warned of a potential uplift enabling real-world cyberattacks; subsequent safeguard updates will apply only to its managed API, not already downloaded copies. Hugging Face hosts the weights under a user-controlled deployment, while Mistral offers a vetted security-focused variant to select partners and plans ongoing guardrail improvements.
