Mistral AI’s new model aims to make content moderation a lot more flexible.
The French AI company on Tuesday released Shieldstral, a 3 billion-parameter multimodal safety classifier that evaluates text and images against moderation policies written in plain language at inference time.
Instead of relying on a fixed set of harm categories built into the model during training, developers can describe their own policy in a yes-or-no question and receive a calibrated safety score without retraining the model. The weights are available under the Apache 2.0 open-source license.
Mistral said Shieldstral can run on a single 16GB NVIDIA GPU and is designed to handle prompt classification, response moderation, refusal detection, and image safety through the same interface.
How it works
Shieldstral treats moderation as a binary question-answering task. Each request includes an instruction that defines the evaluation context and strictness, a yes-or-no policy question, and the content being evaluated, which can be text, an image or a combination of both.
For example, one application could ask whether content provides instructions for exploiting a vulnerability, while another could apply stricter rules around self-harm content. Both would use the same model checkpoint, with the policy supplied at runtime.
According to Mistral, the model returns a continuous probability score based on the likelihood of “yes” or “no,” allowing developers to choose their own enforcement thresholds.
Performance claims
Mistral reported that Shieldstral matches or outperforms open safety models up to seven times its size across several benchmark categories, including text safety, refusal detection, policy adaptability and multimodal moderation.
The company said the model achieved an average F1 score of 84.9% on text safety benchmarks and 83.8% on multimodal safety benchmarks. These figures come from Mistral’s own evaluations across multiple public benchmarks and held-out test sets.
The model was trained on roughly 54.1 million samples, including open-source text data, synthetic contrastive examples and multimodal datasets.
Shieldstral is Mistral’s first moderation model released as open weights. Earlier moderation products from the company were hosted APIs with fixed taxonomies, while Shieldstral allows developers to define policies directly in the prompt and deploy the model on their own infrastructure.
Mistral released the model as part of its participation in the Open Secure AI Alliance, a coalition that includes NVIDIA and other organizations focused on open AI safety and security tools.
The bigger picture
The most interesting part of Shieldstral is not just its size; it is the shift from category-based moderation to policy-based moderation.
In practice, different AI products often need different safety standards, and retraining a moderation model for every deployment is expensive. A lightweight classifier that can adapt through prompts could reduce both cost and engineering overhead for companies building AI products.
The tradeoff is that Shieldstral’s reported performance is based on vendor-authored benchmarks, and Mistral has acknowledged limitations in multilingual coverage and some lower-resource language evaluations. Teams adopting the model will still need to test it against their own users, languages, and risk tolerance.
Related reading: For another glimpse of AI pushing beyond routine tasks, read how OpenAI says its unreleased Astra model advanced 10 long-standing math problems.