Tencent's Youtu-Parsing-Omni 5B Hits 96.96 on OmniDocBench

Tencent has released Youtu-Parsing-Omni, a 5B-parameter omni-modal parser that converts documents, images, audio and video into one JSON schema and scores 96.96 on OmniDocBench v1.6, which the arXiv paper calls state-of-the-art.

The authors describe it as "an omni-modal parser that converts documents, images, audio, and video into structured records grounded in space and time, providing a perception layer for agents that read, locate, and cite evidence across modalities." The design rests on a single Youtu-Omni-Encoder initialized from a pre-trained text language model, with fusion layers that let video frames attend to concurrent audio, replacing the two-tower architectures typical in multimodal parsers.

On OmniParsingBench the paper reports a 75.08 average and says the model "leads all open-weight models on OmniParsingBench," with its advantage concentrated on audio and video modalities. Chart parsing hits 95.02. Model weights are up on Hugging Face, alongside a vLLM plugin and pinned serving settings. It joins a steady run of open-weight multimodal launches our multimodal tracker has logged this quarter.

The release publishes no latency or throughput numbers for the vLLM configuration, and the OmniParsingBench ranking has not yet been independently verified.