Shieldstral: Mistral's 3B open safety guard runs on one 16GB GPU
Apache-2.0 weights match guard models 7x their size, take plain-language policies at inference, and add free content moderation to any app — no retraining.
Copy markdown
A 3B guard that punches 7x its weight
Shieldstral 1.0 matches or beats open safety models up to 7x larger — 84.9% F1 on text safety (level with the 20B GPT-OSS-Safeguard) and 83.8% on multimodal versus OmniGuard-7B's 77.6%. It returns a calibrated score in a single forward pass on one 16GB GPU.
Write the policy as a question, skip the retrain
Instead of harm categories baked into the weights, you hand it a plain-language yes/no policy at inference and get a probability back. Swap or A/B moderation rules per tenant without touching the model — it hits 91.3% F1 on policy adaptability.
Apache 2.0, and it drops into your stack today
Open weights are on Hugging Face (mistralai/Shieldstral-1.0-3B) under Apache 2.0 for commercial use, with day-one vLLM, llama.cpp/GGUF, SGLang and Transformers support. That's a free, self-hosted moderation layer for chat, agent outputs and user images — no per-call API bill.
Know the blind spots before you ship
It covers text and images via a Pixtral encoder across 12 languages, but Mistral flags weaker results on Arabic, Indonesian and deliberately obfuscated inputs. Treat it as a strong first filter, not your last line of defense.