This repository contains the model weights for our paper: "Byte Latent Transformer: Patches Scale Better Than Tokens"

We introduce the Byte Latent Transformer architecture (BLTs), a new byte-level LLM architecture that for the first time, matches tokenization-based LLM performance at scale, with significant improvements in inference efficiency and robustness. BLT encodes bytes into dynamically sized patches, which serve as the primary units of computation. Patches are segmented dynamically based on the entropy of the next byte, allocating more compute and model capacity where there is more data complexity. The BLT architecture includes new attention mechanisms to maximize the information flow between byte and patch hidden representations and a new type of byte-sequence memory. We present the first scaling study of byte-level models up to 8B parameters and 8T training bytes, showing for the first time that we can train a model end-to-end at scale from bytes with no tokenization or other preprocessing. Scaling trends reveal training and inference efficiency benefits from dynamically selecting very long patches on average, along with qualitative improvements with reasoning and long tail generalization from modeling byte-sequences.

To run the model, see the readme here: https://github.com/facebookresearch/blt