Mistral published Shieldstral on August 4, 2026, an open-weights multimodal safety classifier 1. Rather than baking the rule into the weights, it takes the rule in ordinary prose each time it runs. A call comes in three parts: <Instruct> sets the situation and how strict to be, <Query> poses one yes-or-no question such as whether the content promotes physical violence, and <Document> carries whatever is being judged. What comes back is a calibrated probability score, and text and images go through the same path 1.
The reasoning Mistral gives is that conventional guardrails keep their harm taxonomy in the weights, which forces a fresh round of training whenever the deployment changes. What counts as harmful depends on the setting — the company contrasts a security research tool with a mental-health service 1. Shieldstral’s approach is to absorb that difference in the prose handed over at run time. Because scoring reads only the yes and no outputs, sorting inputs, checking responses, and detecting refusals all fit one shape 1.
The model is 3B parameters and fits on a single 16GB GPU 1. That it matches or beats open guard models up to seven times larger on text safety, refusal detection, policy adaptability, and multimodal benchmarks is Mistral’s own claim; the company says evaluation samples were kept out of training 1. The license is Apache 2.0, and the release comes as part of Mistral’s role as an inaugural member of the Open Secure AI Alliance, which it joins alongside NVIDIA and others 1.
Sources
- Introducing Shieldstral. - Mistral AI official blog (August 4, 2026)