← Back to feed

Mistral has released Shieldstral, a 3 billion parameter open-weight multimodal safety classifier that outperforms models up to seven times its size on text safety benchmarks and sets a new state of the art on multimodal moderation. It ships under Apache 2.0 and runs on a single 16GB NVIDIA GPU.

The model rethinks how safety classification works. Instead of baking a fixed taxonomy of harm categories into the weights, Shieldstral accepts a plain-language policy question at inference time and returns a calibrated safety score from a single token. The effect: one model serves any deployment context without retraining.

How It Works

Every moderation request passes three inputs: an evaluation context, a strictness level, and an optional definition of what counts as unsafe content. The model then returns a binary verdict. The same content that passes a cybersecurity research tool can fail a mental-health platform, depending on the policy question asked.

That flexibility matters because harm definitions vary by product, audience, and jurisdiction. Traditional guardrail models encode a fixed view of harm into their weights. Swapping that view requires a new fine-tune. Shieldstral makes the policy a runtime variable.

The model covers both text-to-text and image-to-text evaluation in a unified interface. Most existing safety classifiers treat those as separate pipelines requiring different models.

Benchmark Position

Mistral claims Shieldstral matches models up to 7x its size on text safety benchmarks — roughly 21B-class models — while setting a new SOTA on multimodal moderation tasks. The technical report (arxiv: 2607.25857) provides the full benchmark breakdown across diverse harm categories.

Running a full inference pass requires only a 16GB GPU, which puts it within reach of single-instance deployment rather than a separate safety-model cluster.

Release

Shieldstral 1.0 3B is available on Hugging Face at mistralai/Shieldstral-1.0-3B under Apache 2.0. Mistral is releasing it as an inaugural member of the Open Secure AI Alliance, a joint initiative with NVIDIA and other organizations focused on open-weight safety infrastructure.

The model arrives as content moderation becomes a structural requirement rather than an afterthought. Platforms shipping frontier models face regulatory pressure from the EU AI Act and US state-level AI liability bills. A policy-adaptive classifier that doesn’t require retraining cuts the operational cost of compliance significantly.

Key Numbers

• Parameters: 3B
• Effective benchmark performance: matches 21B-class rivals on text safety
• Multimodal moderation: new SOTA
• License: Apache 2.0
• GPU requirement: single 16GB NVIDIA GPU
• Available: Hugging Face (mistralai/Shieldstral-1.0-3B)