Youtu-Parsing-Omni
An open-source multimodal document parsing model from Tencent with 5B parameters (BF16), built on the Youtu-Omni-Encoder and a Youtu-LLM decoder, supporting parsing of PDFs, tables, charts, images, audio and video and producing structured JSON output.
Timeline (last 90 days)
Tencent has open-sourced Youtu-Parsing-Omni, a 5B-parameter multimodal parsing model that can convert PDF page images, tables, charts, and scans into structured JSON.
Youtu-Parsing-Omni scored 96.96 on OmniDocBench v1.6, ranking first.
The Youtu-Parsing-Omni model has 5.33B parameters and about 10.7 GB of weights (BF16).
The Youtu-Parsing-Omni architecture consists of a shared Youtu-Omni-Encoder feeding into a Youtu-LLM decoder.
Youtu-Parsing-Omni is released in the form of custom code (tag youtu_vita), and calling it requires trust_remote_code=True.
The previous Youtu-Parsing was a 2.5B-parameter model that scored 93.74 on the same leaderboard and could only parse documents; the Omni version adds support for images, charts, audio and video.
Youtu-Parsing-Omni selects modes via the --task parameter, supporting tasks such as document, graphics_chart, graphics_flowchart, graphics_geometric, natural_image, audio, natural_video, textrich_video.
Youtu-Parsing-Omni returns one fenced JSON block per task, containing modality, subtype, elements (or segments) and global_description, with coordinates expressed as tokens on a 0–1000 grid.
On OmniDocBench v1.6, TeleOCR outperforms Youtu-Parsing-Omni on text edit distance and table TEDS; OvisOCR2 and PaddleOCR-VL surpass Youtu on formulas with fewer than 1B parameters.
Youtu-Parsing-Omni has the best reading order score in the table at 0.1104, and its structural table score (TEDS-S) of 98.20 is also first.
5 more timeline events
All Facts (15)
On the OmniDocBench v1.6 leaderboard, Marker 1.8.2 scores 78.44, far below Youtu-Parsing-Omni's 96.96, but Marker can run on a laptop CPU, whereas Youtu-Parsing-Omni requires a CUDA GPU.
50%UnverifiedYoutu-Parsing-Omni scored 96.96 on OmniDocBench v1.6, ranking first.
50%UnverifiedThe Youtu-Parsing-Omni model has 5.33B parameters and about 10.7 GB of weights (BF16).
50%UnverifiedThe Youtu-Parsing-Omni architecture consists of a shared Youtu-Omni-Encoder feeding into a Youtu-LLM decoder.
50%UnverifiedYoutu-Parsing-Omni is released in the form of custom code (tag youtu_vita), and calling it requires trust_remote_code=True.
50%UnverifiedThe previous Youtu-Parsing was a 2.5B-parameter model that scored 93.74 on the same leaderboard and could only parse documents; the Omni version adds support for images, charts, audio and video.
50%UnverifiedYoutu-Parsing-Omni selects modes via the --task parameter, supporting tasks such as document, graphics_chart, graphics_flowchart, graphics_geometric, natural_image, audio, natural_video, textrich_video.
50%UnverifiedYoutu-Parsing-Omni returns one fenced JSON block per task, containing modality, subtype, elements (or segments) and global_description, with coordinates expressed as tokens on a 0–1000 grid.
50%UnverifiedOn OmniDocBench v1.6, TeleOCR outperforms Youtu-Parsing-Omni on text edit distance and table TEDS; OvisOCR2 and PaddleOCR-VL surpass Youtu on formulas with fewer than 1B parameters.
50%UnverifiedYoutu-Parsing-Omni has the best reading order score in the table at 0.1104, and its structural table score (TEDS-S) of 98.20 is also first.
50%UnverifiedYoutu-Parsing-Omni has not yet entered the official OmniDocBench leaderboard; its 96.96 score comes from the Tencent model card, while the older Youtu-Parsing's 93.74 score is on the official leaderboard.
50%UnverifiedOn OmniParsingBench Youtu averages 75.08, below Gemini-3-Pro's 77.44; it wins on charts at 95.02 and audio at 78.08, but clearly lags on natural images at 62.84 versus 69.73 and geometry at 74.33 versus 85.43.
50%UnverifiedThe old model's advertised 5–11x Token Parallelism speedup has not been integrated into vLLM (confirmed in GitHub issue #12), so do not expect that speedup when deploying via vLLM.
50%UnverifiedYoutu-Parsing-Omni uses a custom youtu-parsing license (not Apache or MIT), with a perpetual, worldwide, royalty-free grant and no MAU threshold, but Section 0 excludes the European Union.
50%UnverifiedTencent has open-sourced Youtu-Parsing-Omni, a 5B-parameter multimodal parsing model that can convert PDF page images, tables, charts, and scans into structured JSON.
50%