Shieldstral-1.0-3B

What Is Shieldstral-1.0-3B?

Shieldstral-1.0-3B is a compact, open-weight multimodal safety classifier developed by Mistral AI. Rather than enforcing a fixed set of content categories, it evaluates content against safety policies expressed in plain natural language — covering both text and images — and returns a calibrated safety score from a single model pass.

What It's Good For

Shieldstral is purpose-built for content moderation pipelines in AI products. Its core innovation is policy-adaptive moderation: instead of being locked to one taxonomy, you supply a natural-language yes/no question at inference time (e.g., "Does this content promote physical violence?"), and the model returns a score reflecting compliance with that policy — no retraining required.

Concrete use cases include:

• Prompt classification: Detecting unsafe or adversarial inputs before they reach a main model.
• Response moderation: Checking AI-generated outputs against content policies before delivery to users.
• Refusal detection: Identifying whether an assistant correctly refused a problematic request.
• Toxicity detection: Flagging hate speech, violent content, sexual content, or self-harm references across text and images.
• Multimodal pipelines: One shared interface moderates text-only, image-only, and combined text+image content.

Shieldstral frames every request as three structured parts: an evaluation context with strictness level, a single yes/no query, and the document (prompt, response, image, or pair) to judge. This unified format means a single checkpoint handles novel policies simply by changing the query text.

Because it emits a continuous safety score (not just a binary label), teams can threshold by confidence or rank flagged content — useful for applications that need nuanced filtering rather than hard cutoffs.

Who It's a Good Fit For

Shieldstral is aimed at developers and ML engineers who need to add moderation capabilities to their AI applications without operating a heavy, specialized guardrail system. It fits in 16GB of VRAM (BF16), making it deployable on a single mid-range GPU alongside a main model. It supports inference via vLLM, SGLang, llama.cpp, and the Hugging Face Transformers library.

Teams building on Mistral Large or other Mistral-family models will find Shieldstral naturally complementary. It is also a strong fit for any organization deploying open-weight LLMs via Hugging Face who want an on-premises content safety layer.

Being released under the Apache 2.0 license, it is free to use, modify, and deploy commercially without royalty obligations.

Limitations

• One policy per query: Shieldstral answers a single yes/no question per call; checking multiple independent policies requires issuing one query per policy, which increases latency and cost in high-throughput settings.
• Classifier, not a general model: It is not an instruction-following or reasoning model — its output is always a safety verdict, not a response or explanation.
• Multilingual and long-document coverage: Mistral AI has noted ongoing work on multilingual robustness and handling of longer documents, indicating these areas may perform less consistently than short English-language inputs.
• Self-hosting required for full control: There is no dedicated managed API endpoint specific to Shieldstral listed on Mistral's pricing page; production use at scale may require self-hosting or integration with a third-party inference provider.