Mistral AI releases 3B multimodal safety classifier Shieldstral 1.0
Mistral AI has released Shieldstral 1.0 3B, an open-weights multimodal safety classifier that lets operators define moderation policies at inference time. Instead of relying on a fixed list of harm categories, the system evaluates a plain-language yes-or-no question and produces a calibrated safety score in one forward pass. It is built on Ministral-3-3B-Base-2512 and includes a native Pixtral vision encoder, under the Apache 2.0 license.
Mistral reports an average F1 score of 84.9% for text safety and 83.8% for multimodal safety. The model fits into 16GB of VRAM in BF16 and can run on one GPU. Support is available through vLLM, llama.cpp, SGLang and Transformers, while Axolotl supports fine-tuning.
The classifier uses about 54.1 million samples, including text, synthetic contrastive text and multimodal data. Its continuous output can be thresholded at 0.5 or used to send borderline cases for human review. Mistral notes weaker results for Arabic and Indonesian prompts, adversarial or obfuscated inputs, and very long documents. The trained context length is 32k tokens across 12 languages.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.