7.5
Shieldstral Review — Mistral's 3B Open-Weights Multimodal Moderation Model That Beats Models 7x Its Size
Mistral's Shieldstral-1.0-3B reframes content moderation as policy-adaptive question answering: you write the policy as a plain-language yes/no question at inference time, and a 3B model — built on Ministral-3-3B with a Pixtral vision encoder — returns a calibrated safety score for text, images, or both. Apache 2.0, runs on a single 16GB GPU, matches or beats guard models up to 7x its size (WildGuardTest 88.1, HarmBench 99.4, VLGuard 97.7 F1). Full review with benchmarks, the HN debate on reasoning traces and policy flexibility, and real deployment patterns.