Fast Byte Latent Transformer

We introduce the proposed BLT Diffusion (BLT-D) architecture and related optimization techniques to improve the slow generation speed of the byte-level language model. The existing model maximized reasoning efficiency by introducing Block-wise Diffusion technology to predict multiple bytes in parallel at the same time, away from the existing model generated bytes sequentially one by one. In addition to speed-oriented BLT-D, the researchers propose a self-speculation-based BLT-S and a BLT-DV technique that adds verification steps to supplement performance. This approach has resulted in reducing memory bandwidth costs by more than 50% while maintaining a performance gap with existing tokenization-based models. As a result, these technologies have proven that fast and efficient language models can be implemented without a torque naizer, greatly increasing the practicality of byte-level models. https://arxiv.org/pdf/2605.08044