Mistral AI Releases Shieldstral 1.0 3B: An Open-Weights Policy-Adaptive Multimodal Safety Classifier Matching Models 7× Its Size

Mistral AI has launched Shieldstral 1.0 3B, an open-weights, policy-adaptive multimodal safety classifier that reframes content moderation as a single, binary question rather than applying a fixed taxonomy of harm categories. Unlike conventional guardrail models, which bake their category lists into their weights and require retraining to adapt to new deployment contexts, Shieldstral lets operators define their own policy at inference time. The model then delivers a calibrated safety score in a single forward pass, making it versatile across environments—where content deemed suitable for a cybersecurity research tool might be harmful on a mental-health platform. Built on the Ministral-3-3B-Base-2512 architecture with a native Pixtral vision encoder, Shieldstral is released under the permissive Apache 2.0 license. Its performance is notable: it achieves an average F1 score of 84.9% on text safety tasks, matching the much larger GPT-OSS-Safeguard-20B model, and reaches 83.8% on multimodal safety, outperforming all baselines evaluated by Mistral. ## Is It Deployable? Yes, and it runs locally. The Shieldstral-1.0-3B model fits within 16GB of VRAM when running in BF16 precision, operates on a single GPU, and is licensed for both commercial and non-commercial use. Serving integration is already available through vLLM (version 0.26.0 or later), with additional deployment paths in development.

via MarkTechPost