Access FZ-Temporal-Decision-VisionAI-2B

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

These weights are released under CC BY-NC 4.0 for research and evaluation. Commercial use requires a separate agreement with Fizibox (thanh@fizibox.com). No patent rights are granted. The model is a verification component, not a certified fire alarm. By requesting access you also agree not to use the model, its exports or its outputs (scores, probabilities, labels) to train, distill, fine-tune or otherwise improve any other model.

Log in or Sign Up to review the conditions and access this model content.

FZ-Temporal-Decision-VisionAI-2B (Fizibox)

Developed by the Fizibox Research Team, Vietnam.

A 14-layer Qwen2-VL-2B student that verifies fire and smoke on CCTV by looking at two moments at once: the crop of the current frame and the same crop one second earlier, packed into the temporal patch slot of the vision encoder (no extra vision compute). It answers fixed multiple-choice questions with a single letter.

Question Options
verdict_strict fire / smoke / safe (ignoring named look-alikes)
motion yes / no: does any smoke or flame move, flicker or grow between the two moments?

The model returns calibrated probabilities for each question. Turning them into alarms (which image regions to query, how verdict and motion are combined, temporal confirmation) is part of the Fizibox product and is not included here.

Access is gated: accept the terms on this page, then huggingface-cli login before downloading.

Results (unseen data)

Transition videos (10 incidents) Real webcams, 47 h
This model in the Fizibox alarm logic (verdict + motion) 10/10 detected, 1 false alarm 0 alarms
Same model, verdict only 10/10, 1 0.3-1.3 alarms/h
Single-frame version 9/10, 1 3.6-3.7 alarms/h
WeVisOne-VL-4B (detection, fire/smoke) 10/10, 4 51-62% of frames positive
Base Qwen2-VL-2B, Intel yes/no recipe (full frame, 3-of-4) 8/10, 0 5.4-43.9% of scans positive

Compared with Intel's published recipe for the same 2B base model (untuned, single frame, yes/no prompt), detection is similar but the look-alikes are not: on a 25-incident compilation it raises 12-16 false alarms (ours 0), and on the lake webcam it is in alarm 41.5% of the day (ours 0 alarms). Full table in the paper (v1.1, Section 5.1).

On a Qualcomm QCS8550 NPU (W8A16 decoder, 224x224 vision input) one image pair with both questions takes a median 274 ms end to end; the quantised model detects 9/10 transition incidents on the board, and a frame-by-frame replay of the 47 h webcam recordings on the device raises 0 alarms.

CPU runtimes (gated): FZ-Temporal-Decision-VisionAI-2B-ONNX-OpenVINO (OpenVINO ~0.2 s, ONNX Runtime int8 ~0.7 s per pair with both questions on a Xeon Gold 6530).

Usage

Input: two RGB crops of the same window, 224x224, earlier frame first. Read the softmax over the option letters at the first generated position and divide the logits by the per-question temperature in calib.json. Reference code (MIT): https://github.com/Fizibox-Corp/fz-temporal-decision-visionai

from PIL import Image
from infer import Verifier  # from the GitHub repository

v = Verifier("Fizibox-Corp/FZ-Temporal-Decision-VisionAI-2B")
print(v(Image.open("earlier.jpg"), Image.open("now.jpg")))

Training data

Fizibox in-house CCTV data, public fire and smoke datasets (D-Fire, FASDD), public look-alike images (thliang01/fireworks-night, OpenRAIL; Aeye-coder/Supervised-Fog-Removal-Dataset, Apache-2.0), labelled with soft labels from a larger vision-language teacher (Qwen3.8-27B, Apache-2.0) and with hand-reviewed video pairs. Use of this model must also respect the use restrictions of the OpenRAIL licence of the fireworks images.

Licence

Weights: CC BY-NC 4.0 (research and non-commercial use). Commercial use requires a licence from Fizibox (thanh@fizibox.com). No patent rights are granted. Built on Qwen2-VL-2B-Instruct (Apache-2.0); see NOTICE. Access terms (accepted at the gate): the model and its outputs (scores, probabilities, labels) must not be used to train, distill, fine-tune or otherwise improve any other model. Inference example and event-scoring code: MIT (https://github.com/Fizibox-Corp/fz-temporal-decision-visionai).

Evaluation data

The evaluation videos come from third-party sources and are not redistributed; per-video results are in the paper.

Limitations

Small evaluation sets (10 + 25 incidents), single annotator. A static look-alike that is the main subject of the scene can still be reported. Very slow, uniform smoke is detected late. Small crops are sensitive to JPEG re-compression (later detection); feed raw crops where possible. Not a certified safety device.

Citation

Thanh Nguyen, Fizibox Research Team, Vietnam. Did It Change? Two-Frame Temporal Verification Makes a 2B Vision-Language Model a Practical Fire and Smoke Alarm on Edge NPUs. Preprint v1.1, 2026. https://doi.org/10.5281/zenodo.23661072 (all versions; v1.1: https://doi.org/10.5281/zenodo.23726336)

Downloads last month
16
Safetensors
Model size
2B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for Fizibox-Corp/FZ-Temporal-Decision-VisionAI-2B

Base model

Qwen/Qwen2-VL-2B
Finetuned
(381)
this model
Quantizations
1 model