Abstract: There are numerous factors that can cause external fires in underground coal mines, and once a fire breaks out, the consequences are severe. Early identification and handling of fire hazards have become effective means of risk control. However, the underground environment is full of interfering factors, and there is currently no publicly available dataset for external fires in underground coal mines. This poses many challenges to the speed and detection accuracy of existing target detection algorithms in the early identification of external fires. To solve the problem of the difficulty in accurately identifying external fires in complex mine environments at an early stage, the VI–YOLO model was constructed by embedding the DCNv3 deformable convolution module into the feature extraction network based on You Only Look Once v8s (YOLOv8s), introducing the RepGFPN reparameterized feature pyramid, and adding a P₂ small target detection layer. A visible light fire detection dataset was simultaneously constructed, which included 6 400 multi-scenario fire images, mine non-fire images, and simulated mine fire images. Through conducting a comparison experiment of the VI–YOLO model with other mainstream single-stage target detection algorithms, the VI–YOLO model was accelerated for inference using OpenVINO and TensorRT, and its feasibility in deployment and application on CPU devices and GPU devices was verified. The model was deployed on the low-performance platform Jetson Nano B01, verifying the feasibility of edge-end deployment. The results show that the VI–YOLO model has improved mAP@0.5 by 2.4% compared to the baseline model, with a recall rate increase of 0.8%, and has better detection capabilities for multi-scale fire features. The mAP@0.5 of the VI–YOLO model reaches 90.1%, significantly surpassing similar single-stage target detection algorithms. The detection speed of the VI–YOLO model is 25 f/s, meeting the real-time detection requirements for fires in the underground visual area. After structural optimization and model accuracy quantification in the OpenVINO framework, the inference time of the half-precision model is 49.6 ms, reducing 301.9 ms, and the detection speed reaches 19 f/s, with a speed increase of approximately 6.3 times. After structural optimization and model accuracy quantification in the TensorRT framework, the inference time of the half-precision model is 4.6 ms, reducing 32.4 ms, and the detection speed is 118 f/s, with a speed increase of 4.7 times. In the Jetson Nano B01 deployment and TensorRT framework, the inference speed of the single-precision model is increased to 12.3 f/s, with a mAP@0.5 of 97.6%, and the inference speed of the half-precision model is increased to 15.2 f/s, with a mAP@0.5 of 97.5%, providing new ideas for the deployment and application of the model on CPU, GPU devices, and low-performance devices.