Shieldstral
Mistral AIMistral AI's 3B-parameter multimodal safety classifier, built on Ministral 3 3B with a Pixtral vision encoder. Rather than predicting a fixed set of moderation categories, it answers one yes-or-no question written in natural language about a prompt, a response, a prompt-response pair or an image, and returns a continuous safety score from a single forward pass. It covers twelve languages and fits on a single 16GB GPU.
A ROOST Model Community partner model. Mistral documents it as Shieldstral 1.0, in public preview.
Openness
3 high confidence- weights
- open(Apache-2.0 safetensors on HF mistralai/Shieldstral-1.0-3B, ungated)
- data
- described(the technical report gives the construction recipe for about 54.1M training samples
- license
- Apache-2.0(OSI)
Shieldstral's weights are downloadable without a gate under Apache-2.0, and its technical report explains how the 54.1M-sample training set was assembled. The training data itself is not released, which keeps it at open weights with the other policy-adaptive classifiers here.
- https://arxiv.org/abs/2607.25857 recorded 2026-10-12
The arXiv abstract says "We present the data construction recipe, covering curation and generation of approximately 54.1M samples and a fine-grained evaluation set to evaluate policy adaptability". It links no dataset release.
- https://huggingface.co/api/models/mistralai/Shieldstral-1.0-3B recorded 2026-10-12
The Hugging Face API record reports gated as false, carries the tag "license:apache-2.0", and lists the weight files "consolidated.safetensors" and "model.safetensors". It declares no dataset tag.
- https://huggingface.co/mistralai/Shieldstral-1.0-3B/raw/main/README.md recorded 2026-10-12
Raw model card front matter declares "license: apache-2.0" and "base_model: mistralai/Ministral-3-3B-Base-2512"; the features list reads "Apache 2.0 License: Open weights for both commercial and non-commercial use", and the License section says "This model is licensed under the Apache 2.0 License". Searched the card for dataset and training data: it links no training set, and mentions the training data only as a limitation ("domains represented unevenly in the training data").
Adoption
2 high confidenceAdoption is measured by Hugging Face downloads of the one checkpoint Mistral publishes, Shieldstral-1.0-3B. Mistral also serves it through its own API, and those calls are not counted.
- https://huggingface.co/api/models/mistralai/Shieldstral-1.0-3B recorded 2026-10-12
37,660 downloads in the trailing 30 days for mistralai/Shieldstral-1.0-3B
Capability
4 medium confidenceLike gpt-oss-safeguard, Shieldstral checks content against a policy the user writes rather than a fixed taxonomy, and on Mistral's own benchmark tables the two trade wins at a seventh of the size. Shieldstral also moderates images, but it answers only yes or no and gives no reasoning, so it sits level with gpt-oss-safeguard rather than ahead.
- https://huggingface.co/mistralai/Shieldstral-1.0-3B/raw/main/README.md recorded 2026-10-12
The card calls Shieldstral a "policy-adaptive multimodal safety classifier" that "evaluates content against a safety policy expressed in natural language and returns a single continuous safety score", lists twelve languages, and says it "fits in 16GB of VRAM in BF16". Its prompt-classification F1 table gives WildGuardTest 88.1 for Shieldstral-3B, 87.3 for GPT-OSS-Safeguard-20B and 88.2 for Qwen3Guard-8B, and its multimodal table gives VLGuard 97.7.
- https://mistral.ai/news/shieldstral recorded 2026-10-12
Mistral's announcement, dated August 4, 2026, describes "a 3B open-weights, policy-adaptive multimodal safety classifier that matches models up to 7x its size on text safety and sets a new state of the art on multimodal moderation".
Verified 2026-10-12, except some scores, still awaiting re-confirmation