Sign in
Pending benchmarksTracking since Nov 8, 2026

Is Youtu-Parsing-Omni state of the art?Pending benchmarks.Announced Oct 10, 2026; no independent results yet.

Tencent Youtu Lab's roughly 5B omni-modal parser converts document pages, images, charts, flowcharts, geometry, audio and video into structured OmniSchema JSON. A shared Youtu-Omni-Encoder handles perception across modalities, with modality-specific stems/projectors and a text decoder. Weights and self-hosting examples are public; no public hosted API is confirmed.

Announced facts

The details so far

Lab
Tencent
Announced
Oct 10, 2026
API
Not yet
Open weights
Published
Parameters
5B advertised; 5.334B safetensors parameters
Inputs and output
Document/image, audio and video inputs produce structured text/JSON with spatial boxes, temporal segments, OCR/ASR, captions and reports under one OmniSchema envelope
Architecture
One shared perception encoder with thin modality-specific stems and merger/projector heads; audio/video alignment occurs inside perception; Youtu-LLM text decoder
License
Custom License Terms of Youtu-Parsing (HF tag other, license_name youtu-parsing). Includes an overriding Additional Territorial Limitation: NOT INTENDED FOR USE WITHIN THE EUROPEAN UNION. Not unmodified Apache-2.0; third-party components retain their own terms
Serving
Self-hosted vLLM plugin or Transformers/custom code; separate environments required by pinned Transformers versions; ffmpeg needed for audio/video. Examples are not evidence of a Tencent-hosted API
Limitations
Autoregressive end-to-end decoding does not implement the parallel acceleration of older Youtu-Parsing. Lab reports remaining weaknesses in ASR, knowledge recognition, counting and geometric relationships
Related
Separate from older 2.5B Youtu-Parsing, Youtu-VL, Youtu-LLM and Logics-Parsing-Omni
Date evidence
Public HF conversion discussion reports attempts on this exact model on Oct 10 ET (Oct 11 00:37 UTC). This establishes public access before the Oct 29 paper; Oct 9 repo creation/upload alone does not establish public availability

Tags marked unsure are not stated on the lab's own pages.

Next steps

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

  • Publicly announced with a source
  • Usable by the public (API or weights)
  • Listed on an independent board we track
  • Enough results for a rank in a category
  • Holds across two snapshots (crowns only)
Self-reported

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: OmniDocBench v1.6 Overall 96.96 and OmniParsingBench average 75.08. The latter uses the lab's own re-scoring toolkit with schema repair and corrected annotations, excluding unavailable/problematic samples; not directly comparable with benchmark-author scores.

Receipts

Sources

REPOTencent Youtu-Parsing-Omni model card and weightshuggingface.co/tencent/Youtu-Parsing-OmniREPOPublic conversion discussion: historical access evidence onlyhuggingface.co/tencent/Youtu-Parsing-Omni/discussions/1 · Oct 10, 2026REPOYoutu-Parsing custom weight/code licensehuggingface.co/tencent/Youtu-Parsing-Omni/blob/main/LICENSELABTencent Youtu-Parsing-Omni technical report v1arxiv.org/abs/2610.39301v1REPOTencent official parsing codegithub.com/TencentCloudADP/youtu-parsing

Tracking continues until an independent board lists it.