Mistral launches Shieldstral, a 3B AI model that runs on a single 16GB GPU
New Delhi: Mistral AI has released Shieldstral, a 3-billion-parameter safety classifier that can check text, images and mixed content against policies written in plain language. The open-weight model is available under the Apache 2.0 licence and can run on a single GPU with 16GB memory.
The French AI company says Shieldstral can adapt to different moderation rules without fresh training. A developer can ask a direct question such as whether an image is safe for children or whether a response promotes violence. The model then returns a score based on “yes” and “no” probabilities.
How Shieldstral checks AI content
Shieldstral treats moderation as a question-and-answer task. Each request contains an instruction, a safety question and the material being checked.
The document can include:
- A user prompt
- An AI-generated response
- A prompt and response together
- An image with optional text
This format lets developers change safety rules through the prompt. A cybersecurity service, for instance, may allow technical discussion that a children’s platform would block.
Mistral said, “The core idea is that a small model can beat much larger ones if the data is right.”
Shieldstral benchmark results
According to Mistral’s tests, Shieldstral recorded an overall F1 score of 84.9 across 13 text-safety benchmarks. This placed it level with GPT-OSS Safeguard 20B and ahead of Qwen3Guard 8B at 84, Nemotron 3.5 Content Safety 4B at 83.3 and LlamaGuard 4 12B at 69.1.
In refusal detection, Shieldstral scored 91.5. GPT-OSS Safeguard 20B led this test with 93.7, followed by Shieldstral, Qwen3Guard 8B at 90.7 and WildGuard 7B at 90.2.
Shieldstral posted 91.3 in policy adaptability, behind GPT-OSS Safeguard’s 94.5 and Nemotron 3.5 Content Safety’s 91.8. It finished ahead of ShieldGemma 9B at 89.3 and LlamaGuard 4 12B at 38.8.
Its strongest lead came in multimodal safety. Shieldstral scored 83.8, ahead of OmniGuard 7B at 77.6, Nemotron 3.5 Content Safety 4B at 73.9 and LlamaGuard 7B at 71.6. Mistral says all evaluation samples were kept separate from the model’s training data.
Why the release matters
Content moderation systems commonly rely on fixed categories built during training. Shieldstral moves those rules into natural-language requests, which could make the same model useful across social platforms, AI assistants and enterprise tools.
The company says it is now working on better multilingual coverage, longer documents and broader image-safety checks.