Mistral’s Shieldstral ties a 20B safety model with just 3B parameters

Mistral’s 3-billion-parameter Shieldstral safety classifier matches OpenAI’s roughly seven-times-larger GPT-OSS-Safeguard-20B on combined text benchmarks and leads several multimodal rivals. Its open-weight design also lets operators define moderation rules in plain language at runtime, without retraining the model. Source

Læs hele artiklen hos kilden.