Vision-language models (VLMs) can reason about scene semantics, reducing false alarms and improving fire detection reliability, yet their large size hinders deployment on embedded sensors. This work investigates compressing domain‑specialized VLMs for fully on‑device fire understanding without sacrificing safety‑critical behavior. We introduce a teacher‑student knowledge distillation pipeline: a large VLM fine‑tuned on fire data serves as the teacher, and its logits and intermediate features guide a lightweight student. Experiments span several VLM families (e.g., Qwen, LLaVA) and model scales, showing that compact students retain most of the teacher’s fire‑understanding capability. The distilled models are then deployed on a commercial Detectium fire sensor, where reasoning accuracy, latency, and memory usage are jointly evaluated. Compression impacts not only overall accuracy but also failure modes; Qwen2.5-0.5B offers the strongest trade‑off among the tested variants. These findings provide practical guidance for deploying domain‑specific VLMs in resource‑constrained, safety‑critical environments.
Review