StarDoc-AI's TeleOCR Beats Gemini 3 Pro at Document Parsing With 1.2B Parameters
A 1.2B parameter open-source vision-language model that parses both digital PDFs and warped phone-camera document photos in one framework
- StarDoc-AI released TeleOCR, a 1.2B parameter open-source VLM for unified document parsing
- Handles both digital PDFs and warped camera-captured pages in a single model, no dewarping needed
- Scores 96.87 on OmniDocBench v1.6, beating Gemini 3 Pro, Qwen3-VL-235B, and GPT-5.2
- Won the ICDAR2026 Sci-ImageMiner competition and topped the EMNLP 2026 Dr.DocBench Challenge
- Apache-2.0 licensed, works with Transformers, vLLM, SGLang, and llama.cpp via GGUF
- Introduces Multi-node Consensus Voting, geometry-aware modeling, and Content-Structure Decoupled Learning
TeleOCR targets PDFs and distorted photos with one 1.2B model
StarDoc-AI has released the TeleOCR model card, a roughly 1.2-billion-parameter vision-language model for document parsing. Built on Qwen2.5-VL and released under Apache-2.0, the model handles born-digital pages, scans, and camera captures within one checkpoint. That design can reduce the number of models required to process warped pages, tables, formulas, and complex layouts.
Camera-document pipelines commonly rectify perspective, detect layout regions, run optical character recognition, and invoke separate table or formula parsers. TeleOCR generates content and page geometry directly from the source image, including multi-point layout polygons that can follow curved or skewed regions. Tables are serialized in Open Table Structure Language, or OTSL, while formulas are returned as LaTeX.
1.2B parameters, stronger reported scores
The project reports leading results on three document-parsing suites, including benchmarks designed around degraded and photographed pages. Scores below use the project’s 100-point presentation, with higher values indicating better performance.
| Benchmark | TeleOCR | Reported comparison | Evaluation focus |
|---|---|---|---|
| OmniDocBench v1.6 | 96.87 overall | Gemini 3 Pro: 92.85 Qwen3-VL-235B: 89.78 GPT-5.2: 86.52 |
Text, layout, tables, and formulas across varied documents |
| OmniDocBench tables | TEDS: 97.05 TEDS-S: 98.52 |
Reported first place | Table content and structural fidelity |
| Wild OmniDocBench | 88.53 overall | Reported leading score | Pages captured under real-world conditions |
| PureDocBench, Real Degraded | 70.85 | MinerU2.5-Pro: 62.56 | Degraded document images |
The OmniDocBench leaderboard also places TeleOCR ahead of specialized models including OvisOCR2, PaddleOCR-VL-1.6, and MinerU2.5-Pro. TEDS stands for Tree Edit Distance-based Similarity, a metric that compares the structure and content of predicted tables with reference tables. TEDS-S concentrates on structural accuracy.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
