Compare
Llama Guard 3 8B vs Llama Guard 4 12B
Put two to four models side by side on the fields that decide a deployment: context, price, region, license, languages, hosting, and what is verified.
| Llama Guard 3 8B | Llama Guard 4 12B | |
|---|---|---|
| Provider | Meta | Meta |
| Parameters | 8B | 12B |
| Context window | Not verified: the provider has not documented this claim. | Not verified: the provider has not documented this claim. |
| License | llama3.1 | other |
| Modalities | text | text, vision |
| Release date | 22 Jul 2024 | 23 Apr 2025 |
| Languages | Strong: en Adequate: — | Strong: en Adequate: — |
| Open weights | yes | yes |
| Tasks | Sensitive data, Customer support | Sensitive data, Vision, Customer support |
Where two models differ, this table shows the difference and stops. LLM EU publishes no winner badge without a documented score for that task, and a missing figure stays missing instead of becoming a zero.
When to use it
Llama Guard 3 8B
Strengths
- open weights are published on Hugging Face and can be downloaded without a vendor agreement
- at 8B parameters it fits on a single accelerator
- the model page lists en
- instruction-tuned checkpoint with a chat template in the repository
Limits
- catalogue only — LLM EU does not run this model, and nothing on this card implies an LLM EU endpoint
- no evaluation has been run by LLM EU
- the published configuration records no context window, so the card carries none
- at 8B parameters, the weights fit on a single modest accelerator
- a classifier, not a chat model: it labels prompts and responses against a safety taxonomy instead of answering them
Tasks
Llama Guard 4 12B
Strengths
- open weights are published on Hugging Face and can be downloaded without a vendor agreement
- at 12B parameters it fits on a single accelerator
- the model page lists en
- instruction-tuned checkpoint with a chat template in the repository
Limits
- catalogue only — LLM EU does not run this model, and nothing on this card implies an LLM EU endpoint
- no evaluation has been run by LLM EU
- the llama4 (other) licence is a custom agreement whose commercial terms were not established from its published text, so commercial serving is not permitted until a human clears it; the router must not route to it
- the published configuration records no context window, so the card carries none
- at 12B parameters, self-hosting needs a single accelerator with enough memory for the weights
- vision checkpoints are heavier to serve than their parameter count alone suggests, because the image encoder and pre-processing run alongside the language model
- a classifier, not a chat model: it labels prompts and responses against a safety taxonomy instead of answering them
Tasks
Tasks: Sensitive data, Customer support