Shieldstral
#8 v Multimodální modelyMistral AI · 8× · naposledy 06. 8. 2026
67
Momentum
Shieldstral is a 3-billion-parameter content moderation model published by Mistral AI, available as an Open-Weight model under the Apache 2.0 license. It builds on Ministral-3B and uses a Pixtral vision encoder for multimodal inputs (text and image). Rather than relying on fixed categories, the model processes moderation policies as freely formulated natural language queries at runtime and delivers a calibrated safety score without requiring Fine-Tuning. According to the technical report, it achieves performance comparable to models up to 7 times larger on text safety Benchmarks and sets new records for multimodal moderation.
Vývoj momenta
08.05.06.08.
Vlastnosti
| Key Benchmark (%) | Average F1 score of 84.9% across 16 safety benchmarks; 83.8% F1 on multimodal safety benchmarks |
| Context Window (Tokens) | 32,000 tokens |
| License | Apache 2.0 (Open Weights) |
| Multimodality | Text and image (prompt, response, prompt-response pair, image with optional text) via a unified interface |
| Platform | Runs on a single 16GB NVIDIA GPU, on-device deployable; weights via Hugging Face |
| Release Date | August 4, 2026 (Public Preview) |