Website · Documentation · GitHub · Model card
GGUF weights of Spellguard-1.0-8B, an open-weights guard model for AI agents and chat applications. It flags harmful prompts, harmful replies, and prompt-injection and jailbreak attempts. Results, training and limitations are on the main model card, and every open-weights score there was measured with this file.
| File | Quantisation | Size |
|---|---|---|
Spellguard-1.0-8B-Q4_K_M.gguf |
Q4_K_M | 5.03 GB |
llama-server -hf Spellguard/Spellguard-1.0-8B-GGUF:Q4_K_M -c 8192 -ngl 99
curl -s http://localhost:8080/v1/chat/completions -H 'Content-Type: application/json' \
-d '{"messages":[{"role":"user","content":"How can I make a bomb?"}],"temperature":0,"max_tokens":64}'
# "content": "Safety: Unsafe\nCategories: Violent"
For reply moderation, render the chat template yourself and call /completion, because the chat endpoint treats a trailing assistant message as text to continue. For the score the published results use, P(Unsafe) + P(Controversial) at the first answer token flagged at 0.5, see llama_server_example.py.
Apache 2.0, derived from Qwen3Guard-Gen-8B (Apache 2.0; see NOTICE).
Spellguard is a control plane that secures how AI agents communicate and act. Inside the platform, Spellguard-1.0-8B works together with Spellguard's other security systems and scores higher on prompt injection and content safety; see the main model card.