DFlash-2 AI Model Aims to Improve Draft Token Prediction Speed and Accuracy

Z-Lab has developed DFlash-2, a successor to its DFlash draft-token prediction technique designed to speed up AI text generation. The original DFlash used a diffusion model to predict upcoming output tokens but often suffered from accuracy loss and inconsistent token ordering. The new version adds a Lightweight Path Selector to filter unnatural token sequences and a Local Convolution layer to address accuracy decay in longer predictions. As of August 2026, only a limited number of AI models support the DFlash-2 technique. The enhancements aim to improve both throughput and accuracy compared to the original implementation.