Mistral has released Shieldstral, an open model with 3 billion parameters for checking content safety
Mistral has released the open-weight model Shieldstral (3 billion parameters) for classifying the safety of text and images. According to the company, it achieves an F1 score of 84.9 % on text (on par with the seven-times-larger GPT-OSS-Safeguard-20B) and 83.8 % on images. Rules can be configured at runtime in natural language without training.
Mistral has released Shieldstral, an open-weight model with 3 billion parameters designed as a guardrail layer for checking the safety of text and image content. The model is based on the Ministral-3B architecture with the Pixtral visual encoder, and Mistral co-founder Guillaume Lample contributed to its development. The model is available under the Apache 2.0 license.
According to the description in the related research paper, Shieldstral does not use a fixed taxonomy of risk categories like conventional guardrail models. Instead, operators provide rules as natural-language questions with yes/no answers (e.g. "Does this content promote violence?") at runtime, without needing to retrain the model. The model calculates a safety score from zero to one based on the probability of the answer. The authors state that public safety datasets categorize risks inconsistently and that the same rules are not suitable for every deployment. The model was trained on synthetic data combining 54.1 million examples focused on safety, harmful content and manipulation, with varying levels of strictness depending on the data type; another language model was used to rewrite safe text into unsafe variants, and training also included similar but distinct categories that the model was meant to learn to distinguish.
According to Mistral, Shieldstral achieves an F1 score of 84.9 % on combined text benchmarks, matching GPT-OSS-Safeguard-20B from OpenAI (approximately seven times as many parameters) and outperforming Qwen3Guard-8B (84.0 %), Nemotron-3.5-Safety-4B (83.3 %) and LlamaGuard-4-12B (69.1 %). On images and image-text combinations, it achieves 83.8 %, which the company says is a new best score in this category, ahead of OmniGuard-7B (77.6 %) and LlavaGuard-7B (71.6 %). On the test of adaptability to new rules, GPT-OSS-Safeguard-20B leads with 94.1 % compared with 91.3 % for Shieldstral; nevertheless, the authors consider Shieldstral more practical because GPT-OSS-Safeguard-20B and Nemotron-3.5-Safety generate long chains of reasoning that increase computational costs, while Shieldstral returns a single word. In a separate validation test, synthetic data for subtly differentiated categories increased the F1 score by 23.3 percentage points, which the authors identify as the main factor behind the ability of the model to adapt to new rules.
Why it matters
The smaller open-source model allows companies to deploy a content safety classifier without relying on significantly larger and more expensive models, while rules can be adjusted for the specific application at runtime without retraining.
Two audiences, two different impacts
What this means
For individuals
Developers building AI applications gain an open-weight tool that lets them adapt content safety checks to their own rules without training their own classifier.
For a business
Companies operating AI products can deploy a smaller and, according to the company, cheaper open-source content safety classifier with rules that can be configured at runtime, reducing filtering costs compared with significantly larger models such as GPT-OSS-Safeguard-20B.
DevelopmentCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.