The field of artificial intelligence is moving towards efficient deployment of models on edge devices, with a focus on reducing latency, energy consumption, and memory requirements. Researchers are exploring innovative techniques such as saturation-aware convolution, hardware-aware compression, and extreme model compression to enable the deployment of AI models on ultra-low-power microcontrollers and other resource-constrained devices. These techniques have shown promising results, with significant reductions in inference time and energy consumption while maintaining accuracy. Noteworthy papers in this area include: Efficient CNN Inference on Ultra-Low-Power MCUs via Saturation-Aware Convolution, which achieves up to 24% inference time saving with zero impact on neural network accuracy. Hardware-Aware YOLO Compression for Low-Power Edge AI on STM32U5 for Weeds Detection in Digital Agriculture, which enables real-time weeds detection with minimal energy consumption. Extreme Model Compression with Structured Sparsity at Low Precision, which achieves a 20x model size reduction while retaining 99% of the original accuracy. SPEED-Q: Staged Processing with Enhanced Distillation towards Efficient Low-bit On-device VLM Quantization, which enables accurate and stable quantization of complex Vision-Language Models. Ultra-Light Test-Time Adaptation for Vision--Language Models, which improves top-1 accuracy and reduces ECE by 20-30% without updating any backbone parameters. FQ-PETR: Fully Quantized Position Embedding Transformation for Multi-View 3D Object Detection, which reduces latency by up to 75% while preserving accuracy.