In September 2026, arXiv published research on the PRICE system, proposing a method for short-term Bitcoin price prediction based on the 4-bit quantized LLaMA-3 8B model. The system integrates components such as parameter-efficient fine-tuning (LoRA), recursive multi-step inference, integer rounding numerical representation, and zero-temperature decoding. Compared with eight base models based on Transformer and time series, PRICE achieved the lowest prediction error in both the validation and test sets. The study indicates that this method performs better or equally well than professional models under non-time-series pre-trained data, and the contribution of each component to accuracy was verified through ablation experiments.