Llama Guard 4
Open Weights
Meta
Meta's natively multimodal safety classifier, a dense 12B model pruned from Llama 4 Scout and trained jointly on text and multiple images. It screens both inputs and outputs of a language model, emitting a safe/unsafe verdict plus the violated categories from a 14-class hazard taxonomy (S1-S14) aligned to the MLCommons hazards standard. It runs on a single 24 GB GPU.
Strengths
- Natively multimodal - classifies mixed text-and-image prompts, not just text
- Screens both prompts and model responses, so one model covers input and output filtering
- Names the violated categories from a 14-class taxonomy (S1-S14), not just a binary verdict
- Aligned to the standardized MLCommons hazards taxonomy
- Dense 12B pruned from Llama 4 Scout, running on a single 24 GB GPU
Caveats
- A classifier, not a chat model - it labels content rather than answering
- Meta's own output-filtering figures are modest: 69% recall at 11% false positives in English
- Multilingual recall drops to 43%, and single-image recall to 41%
- The Llama 4 Community License is not an OSI-approved open-source licence
- No release date is stated on the model card
- Vendor
- Meta
- License
- Llama 4 Community License
- Modalities
- textimage
| Benchmark | Score | Max |
|---|---|---|
| Output filtering F1 (English) | 61.0 | 100.0 |
| Output filtering F1 (multilingual) | 51.0 | 100.0 |
| Output filtering F1 (multi-image) | 52.0 | 100.0 |
Capabilities
Vision
Audio
Video
Tool Use
Resources
Pricing
Open weights under the Llama 4 Community License - self-host at no licence cost. Fits on a single 24 GB GPU.