Shieldstral 1.0

Shieldstral 1.0 is Mistral AI's first published moderation model, released on 4 August 2026 under Apache 2.0. It does not answer questions or write text; it judges content that other models produce or receive. What makes it different from earlier guardrail models is where the rule lives. Systems such as LlamaGuard or ShieldGemma have a fixed taxonomy of harm categories baked into their weights, so a product that needs a different definition of unsafe content has to retrain. Shieldstral instead reads three parts: a context and strictness instruction, a single yes-or-no question written in ordinary language, and the content to judge. The question is the policy. A cybersecurity forum and a mental-health service can run the same checkpoint with different questions and get different verdicts on the same text. The verdict itself is one forward pass. The model reads only the probabilities of yes and no and normalises them into a continuous score between the two, so an operator can set the threshold and rank borderline cases by confidence instead of receiving a bare label. It is built on the Ministral 3 3B base with a Pixtral vision encoder, which lets a single interface cover text, images and text-with-image. The vendor's tables put it level with or ahead of open guard models up to seven times its size: 84.1% F1 on ToxicChat and 99.4% on HarmBench for prompt classification, against 79.8% and 94.5% for a 20-billion-parameter rival. It is not first everywhere — on the multilingual RTP-LX prompt set it trails clearly — and the vendor names multilingual coverage as the next thing to fix. Practical limits are worth reading before deployment. The architecture supports a 256,000-token context, but training used sequences up to 32,000 and the vendor advises staying inside that range. The model runs on a single 16 GB graphics card, which is the point: moderation is a per-request cost, so a guard that needs a server rack is a guard that gets skipped.

▸Videos

No videos yet.

▸Reviews