Superagent Guard is published as open weights under the superagent-ai organization on Hugging Face. Self-host the same models that power the API: inference runs on your infrastructure, no external API calls required, and no prompts leave your network.

Three sizes

Model	Parameters	Use when
superagent-guard-0.6b	0.6B	Ultra-low latency and edge deployment
superagent-guard-1.7b	1.7B	Balanced accuracy and speed; start here
superagent-guard-4b	4B	Maximum detection when false negatives are expensive

Each size also ships a GGUF variant (-gguf suffix) for local runtimes such as llama.cpp.

Typical latency is 50-100ms on your own hardware, which keeps the check inside the request path.

Details

• Base: Qwen3, fine-tuned for security classification
• Output: structured JSON with a pass or block decision, violation types, and CWE codes
• License: CC BY-NC 4.0. Free for non-commercial use; commercial use requires separate licensing from Superagent
• Access: weights are gated on Hugging Face. Accept the terms on the model page to download
• Training data: the superagent-guard dataset is published alongside the models

The model cards include ready-to-run snippets for Transformers, vLLM, and SGLang.

Next steps

• Guardrails overview
• superagent-ai on Hugging Face