Mistral released Shieldstral 1.0 3B, an openly licensed model that checks text and images against a safety policy written in ordinary language and returns a score for whether the content breaches it. The policy can be changed at any time without retraining the model, and it runs on a single GPU. Mistral reports accuracy at or near larger competing safety models on most public benchmarks, and notes reduced reliability on disguised or encoded text and very long documents.
What changed
Open content-moderation models generally classify against a fixed list of harm categories set at training time, and larger checkpoints were needed for comparable accuracy.
What it unlocks
Running a content safety check on text and images against a policy written in plain language, changed at any time without retraining, on a single GPU or a local machine.
- 3B parameters, fits in 16GB of VRAM in BF16
- trained on sequences up to 32k tokens
- 12 languages supported
- VLGuard multimodal F1 97.7% vs 88.5% for OmniGuard-7B
What you need to act on it
- a GPU with roughly 16GB of memory, or a quantized build for local use
- serving through vLLM, llama.cpp, SGLang or Transformers
- one yes/no query per policy
Sources