Youtu-Parsing-Omni Omnimodal Analysis Model: 12GB+ VRAM

Youtu-Parsing-Omni Omnimodal Analysis Model: 12GB+ VRAM

At a Glance

Item Value
Repository tencent/Youtu-Parsing-Omni
Publisher guide Tencent: models and licenses
Published 2026-10-09
License youtu-parsing
Formats safetensors
Paper arXiv:2601.20430, arXiv:2601.19798, arXiv:2512.24618
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

Tencent has released “Youtu-Parsing-Omni", a multimodal model capable of analyzing diverse media such as documents, images, audio, and video, and outputting the results in a structured JSON format. Despite its compact parameter size of 5B, it possesses “omnimodal" performance, capable of processing both visual and audio information within a single model to comprehensively extract layout elements, text, tables, mathematical formulas, timestamps, and audio events.

According to the accompanying paper, “Youtu-Parsing: Perception, Structuring and Recognition via High-Parallelism Decoding", the model was developed with the aim of achieving high-accuracy content extraction while delivering a 5x to 11x speedup compared to conventional autoregressive decoding by adopting a high-parallelism decoding strategy. This model aims to strike a balance between practical utility and experimental value in the field of document intelligence.

Specifications

The specifications listed in the documentation are as follows:

  • Parameters: 5B
  • Architecture: YoutuVITAForCausalLM (youtu_vita)
  • Context Length: 65,536

Performance

The benchmark scores on “OmniDocBench" published by the creators are summarized in the table below, along with major comparison targets. Due to the wide variety of tables, some models and items have been excerpted.

→ Scroll horizontally to see all columns

Model Size Overall ↑ Text Edit ↓ Formula CDM ↑ Table TEDS ↑
Youtu-Parsing-Omni 5.0B 96.96 0.0271 96.80 96.79
TeleOCR 1.2B 96.91 0.0267 96.59 96.82
Gemini 3 Pro – 92.91 0.0640 95.99 89.15
Qwen3-VL-235B 235B 89.78 0.0630 92.55 83.07
GPT-5.2 – 86.59 0.1140 88.21 82.95

From this comparison table published on the model card, it can be seen that Youtu-Parsing-Omni recorded an overall score of 96.96, representing the highest value among the comparison targets. It is particularly noteworthy that it outperforms major models such as Qwen3-VL with its massive 235B parameter count, GPT-5.2, and Gemini 3 Pro in the field of document analysis.

Looking at detailed items, although it falls slightly behind TeleOCR in Table TEDS (which indicates table structure recognition accuracy), it maintains an extremely high standard in overall analysis capability. Furthermore, according to the complete tables in the documentation, it also achieves top-class results in Reading Order Edit (0.1104), which indicates reading order accuracy, and Table TEDS-S (98.20), which indicates table structure comprehension capability. On the other hand, in Text Edit (which indicates text edit distance) and Formula CDM (formula recognition), it occasionally trails slightly behind lightweight models specialized in specific domains such as OvisOCR2 and PaddleOCR-VL-1.6, but overall it can be said to be a highly balanced and high-performing model.

Strengths and Use Cases

The greatest strength of Youtu-Parsing-Omni is its “omnimodal" analysis capability, which allows it to process different modalities—documents, images, audio, and video—within a single model and output them in a consistent JSON format. By switching the task prompts (--task) described in the model card, it can handle an extremely wide range of applications as follows:

Document and Static Image Analysis

In document analysis (document), it can output not only text and layout element extraction, but also complex mathematical formulas in LaTeX format, tables and graphs in Markdown format, and even flowcharts in Mermaid format as structured data. For natural images (natural_image), in addition to detecting entities and texts within the image, it is capable of generating detailed captions and narratives. Furthermore, in the analysis of geometric figures (graphics_geometric), it can extract not only shapes such as points, lines, and arcs, but also their geometric relationships and measurements, making it suitable for digitizing technical drawings.

Multimodal Audio and Video Analysis

In audio analysis (audio), it goes beyond speech-to-text transcription (ASR) to perform speech and non-speech segment division, timestamp assignment, speaker identification, and even caption generation for timbres, surrounding scenes, and acoustic events. In video analysis (natural_video, textrich_video), it analyzes visual actions, interactions between objects, and camera movements, while also integrating the contents of the audio track. Especially for text-rich videos like lecture videos, it can create detailed structured reports combining OCR and ASR, making it powerful for summarizing long videos and creating indexes.

High-Speed Processing and Practicality

According to the paper “Youtu-Parsing: Perception, Structuring and Recognition via High-Parallelism Decoding", the model adopts unique decoding strategies called “token parallelism" and “query parallelism". In token parallelism, up to 64 candidate tokens are generated simultaneously in a single inference step and verified, achieving a 5x to 11x speedup compared to conventional autoregressive decoding. This is particularly effective in generating structured data such as table recognition. Additionally, query parallelism allows content prediction for up to 5 regions (bounding boxes) to be executed simultaneously, enabling an additional 2x speedup without compromising quality. Furthermore, it is reported to possess strong robustness against multilingual text, handwritten characters, and rare characters.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 5.3B parameters

Your VRAM Quantization File size Est. memory needed
12GB (RTX 4070 / 3060 12GB, etc.) BF16 9.9GB 11.9GB

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-11-17): llama.cpp: not registered, vLLM: not registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Can You Run It Locally?

Not usable in Ollama, LM Studio and llama.cpp yet.

The publisher ships safetensors only, and llama.cpp’s registry does not list this architecture. llama.cpp would need to add support before these tools can run it. 2 converted build(s) from other uploaders exist. Today it can be run with transformers, using the memory figures in the table above.

License — youtu-parsing: A custom license from the publisher. Check the original terms directly, including whether commercial use is permitted.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Measurements We Did Not Take

We have not confirmed that this site meets the commercial-use terms of this model’s license (youtu-parsing). Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).

Recent Models in the Same Size Class

Models with 4–15B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

→ Scroll horizontally to see all columns

Model Parameters Smallest VRAM tier License Our article
nvidia/Muse-Spark-Open-816B-A42B-DFlash2 4.4B 12GB nvidia-open-model-agreement-license Muse-Spark-Open-816B-A42B-DFlash2 Vision-Language Model: 12GB+ VRAM (2026-11-17)
tencent/WeVisOne-VL-4B 4.8B 4GB apache-2.0 Tencent Releases WeVisOne-VL Vision Language Models (2026-11-17)
inclusionAI/ZwZ2-9B-GGUF 9.4B 8GB apache-2.0 ZwZ2-9B-GGUF Vision-Language Model: 8GB+ VRAM (2026-11-17)
inclusionAI/ZwZ2-4B-GGUF 4.5B 4GB apache-2.0 ZwZ2-4B-GGUF Vision-Language Model: 4GB+ VRAM (2026-11-17)
microsoft/Mage-VL-Cowork 4.7B 12GB apache-2.0 Mage-VL-Cowork Vision-Language Model: 12GB+ VRAM, GGUF Builds (2026-11-17)

How to Get It

The repository for Youtu-Parsing-Omni is available on Hugging Face. The distribution format is safetensors, and it is not a gated model requiring license agreement.

Using Inference Engines

This model can be run using vLLM or Transformers. The creators recommend operating it with vLLM to obtain reproducible inference results, and provide a dedicated plugin vllm-plugin-vita-omni.

Launching the server using vLLM is performed as follows using the included script:

MODEL=tencent/Youtu-Parsing-Omni DATA_ROOT=/path/to/media GPU=0 bash scripts/vllm.sh

This server provides an OpenAI-compatible API and accepts requests at http://127.0.0.1:8000/v1.

Using with Transformers

When using the transformers library directly, version 5.10.2 or higher is required. When loading the model, it is necessary to specify trust_remote_code=True to allow execution of the custom code included in the repository. The measurement results provided by the creators were calculated using vLLM.

Request Notes

When executing inference, it is necessary to select and send an appropriate task prompt from prompts/youtu_parsing_omni.json included with the model. Additionally, coordinate information is output as proprietary tokens based on a 0-1000 grid, such as <box><x_85><y_275>..., and the HH:MM:SS format is used for timestamps. When making API requests, it is necessary to set enable_parsing: true as a chat template argument and configure it to retrieve responses without skipping special tokens.

Related Articles

What to Read Next

Sources