Back to Model
D

Decima Guard

Model Oct 18, 2026

A community 321M CPU-oriented classifier preview for prompt injection, tool-result inspection and output checks in English, Arabic and Persian.

Visit primary link 2 sources attached
Decima Guard's author-published terminal demo flags an instruction to reveal a system prompt as prompt injection.
Full frame from the author's English CLI animation. The scores and 1605 ms shown belong to this demonstration, not an AITrending test or the separate per-question speed benchmark. Source: A. M. Madani / Decima
self-hostedverification

What happened lately

Decima Guard timeline

Decima Guard is a community classifier with 321 million parameters and a 2,048-token input limit. Its October 18 preview provides int8 and FP32 ONNX files and PyTorch weights. The English model card documents installation through decima-ai and presets for prompts, retrieved material, model output and web inputs.

The author reports local CPU inference and multilingual evaluation, including fewer benign-message false alarms than a DeBERTa baseline on one comparison. Attack misses remain material. The checkpoint declares Apache-2.0, while its card discloses training sources without published licenses. AITrending reviewed the documentation without reproducing the results; the model is one application layer, not an execution-permission boundary.