Did It Change? Two-Frame Temporal Verification Makes a 2B Vision-Language Model a Practical Fire and Smoke Alarm on Edge NPUs

Vision-language models (VLMs) are increasingly used to verify fire and smoke alarms raised by lightweight detectors, but a single frame cannot separate a fire from the many things that look like one: sunsets, lamps, steam, fog, glare, red or yellow objects. We show that this failure is not solved by scale: two recent open grounding VLMs (WeVisOne-VL 2B and 4B) flag 46-62% of all frames of two ordinary public webcams over 23 hours each. We present FZ-Verdict, a 2B-class VLM student that answers fixed multiple-choice questions with a single letter and looks at two moments at once: the current frame and the frame one second earlier are packed into the temporal patch slot of the Qwen2-VL vision encoder, so the second frame costs no extra vision compute. Besides the verdict question, the model answers a motion question (does any smoke or flame change between the two moments?) trained with counterfactual labels that cannot be solved from the current frame alone. Combining the two answers in our alarm logic, on 47 hours of unseen real-camera footage the system raises zero alarms, versus 3.6-3.7 per hour for our own single-frame model and near-continuous alarms for the grounding VLMs, while detecting 10/10 incidents (median 3.5 s after onset) on a set of unseen transition videos including creeping smoke the detector never boxes. The quantised model runs on a Qualcomm QCS8550 NPU in a median 274 ms per image region with both questions sharing one prefill, and keeps 9/10 detections on the board. Model: https://huggingface.co/Fizibox-Corp/FZ-Temporal-Decision-VisionAI-2B Code: https://github.com/Fizibox-Corp/fz-temporal-decision-visionai