Meta releases Llama Guard 4, a 12B multimodal safety classifier
For context: a dense 12B pruned from Llama 4 Scout that screens prompts and responses across 14 MLCommons hazard categories. Gated under the Llama 4 licence.
For context: Llama Guard 4 is from April 2025, outside the last twelve months, and is included because it defines the category. The card gives the licence effective date as 5 April 2025 and no exact release date. The Hub repository was created on 23 April 2025.
Llama Guard 4 is a natively multimodal safety classifier with 12 billion parameters that handles text and images. It is a dense model pruned from Llama 4 Scout and screens both prompts and LLM responses. It uses 14 hazard categories, S1 to S14, based on the MLCommons taxonomy, from Violent Crimes to Elections. S14, Code Interpreter Abuse, applies to text only.
The licence is the Llama 4 Community License Agreement, with attribution and Built with Llama requirements and an Acceptable Use Policy. Access is gated: you log in, accept the licence and fill a form requiring your full legal name, date of birth and organisation.
On Meta's in-house test set for output filtering, the card reports English recall 69%, false positive rate 11% and F1 61%. Multilingual F1 is 51% and single-image F1 is 38%. Meta says it omitted comparisons with competitor models.
The numbers are modest and the card is plain about it. It stays on the list as context.
Sources
Editor's first pass (AI-assisted)
- Real 4: Weights available after accepting the licence and a gated access form.
- Useful 3.5: Screens text and images in and out; Meta's own table shows modest recall.
- Notable 4: An earlier reference point for open guard models.
- Novel 3: An update to earlier Llama Guard versions, now natively multimodal.
Comments