Primary use stated or inferred cautiously from official documentation.
T
TENCENT / tencent/Youtu-Parsing-Omni
Youtu-Parsing-Omni
Tencent's 5-billion-parameter multimodal model, publicly released on Hugging Face on October 9, 2026. It analyzes pages, images, charts, diagrams, geometry, audio and video, producing structured text in an OmniSchema JSON envelope. It does not generate images, music or video. A shared encoder processes modalities before the decoder; the requested task determines the schema and extracted elements.
01 / SUMMARY
The essential
Maximum input capacity when published by the source.
Documented maximum generation limit.
Verified availability channels.
Primary functional family and verified specialties.
Declared terms for API access, model weights, or self-hosting.
02 / TECHNICAL SHEET
Limits and integration
| API ID | tencent/Youtu-Parsing-Omni |
|---|---|
| Model type | Multimodal · OCR and documents · Voice and audio · Video |
| Access model | Restricted weights |
| License | Youtu-Parsing License · EU territorial restriction |
| Deployment | Self-hosting with vLLM and plugin |
| Release | 09 OCT 2026 |
| Knowledge cutoff | Not published |
| Entrance | Text · Image · Audio · Video |
| Exit | Text · JSON |
| Context window | 65,536 tokens in the serving example |
| maximum output | Not published |
| Reasoning | Not published |
| Published tools | |
| Structured Outings | OmniSchema JSON |
| Batch processing | Not published |
| Prompt cache | Not published |
| Fine-tuning | Not published |
| Verified platforms | Hugging Face · vLLM · Transformers |
OCR, tables, formulas, reading order, geometry, transcription and timestamped audiovisual segmentation. Outputs can include Markdown, LaTeX, Mermaid, boxes and descriptions. The serving example configures 65,536 tokens and limits of eight images, four videos and eight audio inputs per prompt; these are example server settings, not a universal guarantee of the maximum trained context.
The Youtu-Parsing license includes an explicit territorial limitation: the model is not intended for use within the European Union. We do not describe it as Apache 2.0 or unrestricted. This profile is informational: we have neither downloaded nor run its weights. Review the license, remote code, local-media permissions and personal data before considering any authorized deployment.
03 / CAPABILITIES
What it can do
Orientation
OCR, tables, formulas, reading order, geometry, transcription and timestamped audiovisual segmentation. Outputs can include Markdown, LaTeX, Mermaid, boxes and descriptions. The serving example configures 65,536 tokens and limits of eight images, four videos and eight audio inputs per prompt; these are example server settings, not a universal guarantee of the maximum trained context.
Context
65,536 tokens in the serving example · Not published
Tools and integration
Not published
Access
Hugging Face · public repository
Input modalities
TextImageAudioVideo
04 / PRICES
Documented cost
| Concept | Worth | Unit/condition |
|---|---|---|
| Standard input | No published official API rate | Standard API rate |
| Cached input | Not published | Reading reused prefixes |
| Cache write or storage | Not published | The condition varies by provider |
| Standard output | Self-hosted infrastructure cost | May include reasoning tokens |
| Batch input | Not published | Asynchronous processing |
| Batch output | Not published | Asynchronous processing |
No hosted commercial API or public per-token price has been verified for this checkpoint. Numeric cost fields remain empty. Self-hosting requires GPUs, dependencies and the provider's vLLM plugin; publicly available files remove neither costs nor territorial restrictions. Evaluation code is listed as awaiting publication, so full harness reproduction is not promised.
i Prices change and may depend on level, region or context length. Check the source before making a decision.
05 / EVALUATIONS
How to read the results
“
Tencent reports OmniDocBench v1.6 Overall 96.96 and OmniParsingBench average 75.08. It also reports ChemOCR and PDMX-Synth metrics; names and direction are preserved: lower error rates are better. These are provider results, not evaluations run by Inferama. Valid JSON does not ensure correct extraction: inspect reading, coordinates, timing, omissions and hallucinations with human review.
| Benchmark | Result | Metric | Source |
|---|---|---|---|
| OmniDocBench v1.6 Overall | 96,96 | Score | View source ↗ |
| OmniParsingBench Average | 75,08 | Score | View source ↗ |
| ChemOCR Average Similarity | 74,6 | Average similarity | View source ↗ |
| ChemOCR Tani@1.0 | 54,2 | Tani@1.0 | View source ↗ |
| PDMX-Synth CER | 22,73 | CER | View source ↗ |
| PDMX-Synth SER | 26,73 | SER | View source ↗ |
| PDMX-Synth LER | 33,20 | LER | View source ↗ |
Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.
06 / VERSIONS
Chronology
Youtu-Parsing-Omni
Verified public release of the 5B checkpoint. Output is JSON, not generated audiovisual content.
07 / SOURCES