MistralのShieldstral、3BモデルでOpenAIの20Bガードレールを上回る

mistral published shieldstral, a 3-billion-parameter model that matches the performance of safety models nearly seven times its size on standard benchmarks. the model allows operators to define safety rules at runtime using plain-language questions, outputting a single probability score instead of lengthy reasoning chains to reduce compute costs. in tests, shieldstral achieved an f1 score of 84.9% on text tasks and 83.8% on image tasks, tying openai's gpt-oss-safeguard-20b and significantly outperforming qwen3guard-8b and llamaguard-4-12b. mistral co-founder guillaume lample notes that the method addresses the fragmentation of public safety taxonomies and the need for context-aware content moderation.