Tar by ByteDance

Tar by ByteDance is a unified multimodal large language model (MLLM) designed for both image and text understanding and generation. By leveraging a novel text-aligned visual tokenizer (TA-Tok), Tar discretizes images into tokens mapped to an LLM's vocabulary, enabling seamless cross-modal input/output in a shared interface. It supports both high-fidelity visual generation and advanced visual understanding, eliminating modality-specific designs and delivering competitive benchmarks in both speed and accuracy. Tar is aimed at researchers and developers building advanced multimodal AI applications.