Mistral AI has introduced Shieldstral, a three-billion-parameter open-weights model designed to evaluate the safety of text, images and combined text-image content. The company released the classifier under the Apache 2.0 licence and says it can run on a single Nvidia GPU with 16GB of memory.
Shieldstral’s central design choice is to treat moderation as a binary question-answering problem rather than embed a fixed list of harm categories in the model. A request contains an instruction describing the evaluation context and strictness, a yes-or-no policy question, and the material to assess. The document can be a prompt, a response, a prompt-response pair or an image accompanied by optional text.
At inference, the model evaluates the logits for “yes” and “no” and normalizes them into a continuous score. A developer can then select a threshold or use the result for ranking. Because the policy is written as a natural-language query, Mistral says operators can change what the model checks without retraining it. The same checkpoint can be used for tasks including prompt classification, response moderation, refusal detection and toxicity assessment.
Mistral reports that Shieldstral matches or exceeds open guard models as much as seven times larger across its evaluations of text safety, refusal detection, policy adaptability and multimodal moderation. Those results are the developer’s own benchmark claims; the announcement says evaluation samples were held out from training. It also says the model establishes a new high mark on multimodal moderation.
The training process consolidated datasets whose taxonomies and annotation schemes differ. Mistral converted them into a common instruction-query-document structure and varied wording and delimiters to reduce dependence on a single phrasing pattern. The company also generated contrastive examples intended to teach distinctions between similar policies. For visual moderation, it supplemented limited safety datasets with general-purpose images as negative examples and used a vision-language reranker to filter image-query pairs.
Mistral says it fine-tuned separate checkpoints with LoRA, then combined them using spherical linear interpolation. One component targeted calibration on public data, another policy discrimination, while the base instruction model contributed general instruction-following behavior. Training and evaluation ran on Forge, Mistral’s platform for managing infrastructure, sharding, metrics and logs.
The release offers developers a relatively compact, configurable alternative to guardrails built around permanent categories. Mistral says work remains on multilingual support, reliability with longer documents and broader multimodal coverage. Real-world adoption and independent testing will determine how well the company’s reported performance transfers across products and moderation policies. The open licence may also make that scrutiny easier by allowing developers to download and evaluate the weights directly.
