Ministral 3 3B 2512
12B
- Context
- 512tokens
- Released
- 1y ago
- Size
- 86Mparameters
Meta's Llama Prompt Guard 2 86M is an update to the company's Llama Prompt Guard classifier model, built to support the Llama 4 line of models and serve as a drop-in replacement for Prompt Guard. Meta says it improves on the earlier classifier's performance in detecting jailbreak and prompt injection attacks. The BERT-based classifier labels input text as benign or malicious and is trained on both English and non-English attacks to help developers reduce prompt-attack risks in LLM applications.
How it compares
Size class
Tiny86MAn open-weight model under 5 billion parameters, small enough for phones, laptops and edge devices.
Open models of a similar size from other labs
Vs. your stack
Add the models you use to see how this one compares.
Sign inAdditional source evidence
- Source model
- meta/llama-prompt-guard-2-86m
- Source family
- llama
- Last updated
- 2025-04-29
- Model context
- 512
- Model output limit
- 512
- Open weights
- Yes
- File attachments
- No
- Reasoning
- No
- Tool calling
- No
- Temperature control
- No
Downloadable weights
Complete source record (JSON)
{
"id": "meta/llama-prompt-guard-2-86m",
"name": "Llama Prompt Guard 2 86M",
"description": "86M-parameter (mDeBERTa-base) multilingual classifier that detects prompt injection and jailbreak attempts",
"family": "llama",
"last_updated": "2025-04-29",
"attachment": false,
"reasoning": false,
"tool_call": false,
"temperature": false,
"modalities": {
"input": [
"text"
],
"output": [
"text"
]
},
"limit": {
"context": 512,
"output": 512
},
"release_date": "2025-04-29",
"open_weights": true,
"license": "Llama 4 Community License",
"weights": [
{
"label": "Hugging Face",
"url": "https://huggingface.co/meta-llama/Llama-Prompt-Guard-2-86M"
}
]
}Frequently asked questions
Running locally
Estimated VRAM is 0.6 GB at Q4, 0.6 GB at Q8 and 0.7 GB at FP16. All figures are estimates from the 86M parameter count, which is weights plus about 15% and 0.5 GB for a short context.
Yes, it fits at FP16 on a 24 GB, 16 GB, 12 GB or 8 GB GPU, with an estimated 0.7 GB of VRAM.
Yes. Assuming macOS lets the GPU use about 75% of unified memory, the smallest common Mac memory size that fits Q4 is 8 GB. The Q4 VRAM estimate is 0.6 GB.
Privacy and data
Yes, it has open weights, so you can self-host it for full control of your data.
Capabilities and specs
The context window is 512 tokens and the maximum output is 512 tokens.
No, the only input modality is text.
No, tool calling (function calling) is not supported.
No, a reasoning or thinking mode is not supported.
Quality and benchmarks
Closest alternatives are FastVLM 0.5B by Apple, FastVLM 1.5B by Apple and OLMo 2 32B by Ai2, matched on model kind and release date with no comparable price.
FastVLM 0.5BLlama Prompt Guard 2 86M vs FastVLM 0.5BFastVLM 1.5BLlama Prompt Guard 2 86M vs FastVLM 1.5BOLMo 2 32BLlama Prompt Guard 2 86M vs OLMo 2 32B
Versions and lineage
Both have a 512 token context and both were released on April 29, 2025, so the facts show no difference in these specs.
Llama Prompt Guard 2 22MLlama Prompt Guard 2 86M vs Llama Prompt Guard 2 22M
Licensing and access
It is open weights, released under the Llama 4 Community License.
Hugging Face repository: meta-llama/Llama-Prompt-Guard-2-86M.
Basics
Meta released it on April 29, 2025.
Yes, it has open weights.
Generated from Vertical Tabs data, updated November 10, 2026.
Discussion
No comments yet
Be the first to share your thoughts.