# StarDoc-AI Releases TeleOCR for Document Parsing

> A vision-language model built on Qwen2.5-VL targets bilingual Chinese and English text extraction from documents.

Published by The Open Weights on Aug 14, 2026. Canonical: https://theopenweights.com/news/teleocr-huln

## Key facts

- Company: StarDoc AI
- Model: TeleOCR
- Version: 1
- Category: Vision-Language
- Modalities: Vision-Language
- License: Other (Open weights, commercial use allowed)
- Context window: 128K tokens
- Architecture: Qwen2_5_VLForConditionalGeneration
- Precision: BF16
- Weights size: 2.8 GB
- Significance: notable
- Published: 2026-08-14
- Last verified: 2026-09-28
- Hugging Face: https://huggingface.co/StarDoc-AI/TeleOCR
- Canonical URL: https://theopenweights.com/news/teleocr-huln

StarDoc-AI has published [TeleOCR](https://huggingface.co/StarDoc-AI/TeleOCR), a vision-language model aimed squarely at document parsing and optical character recognition. Built on top of Alibaba's Qwen2.5-VL foundation, the model is designed to read and extract text from documents in both Chinese and English.

OCR remains one of the most practical applications of multimodal AI. While general-purpose vision-language models can describe images and answer questions about them, document parsing demands precise handling of dense text, tables, and layout — the kind of structured extraction that businesses rely on for digitizing records, invoices, and forms.

## Why it matters

By fine-tuning an established open model rather than training from scratch, StarDoc-AI inherits Qwen2.5-VL's visual grounding while specializing it for a narrower, higher-value task. Bilingual support is notable here:

- Chinese-language document OCR is underserved by many Western-built models
- English coverage keeps the model useful for cross-border workflows
- A specialized OCR model can outperform generalists on layout-heavy pages

The release arrives under a custom license, and the record does not specify parameter count or context length. Teams evaluating document-processing pipelines will want to test TeleOCR against both dedicated OCR engines and the broader family of open vision-language models before committing. The weights and details are available on the model's [Hugging Face page](https://huggingface.co/StarDoc-AI/TeleOCR).

## Get the model

- [Hugging Face](https://huggingface.co/StarDoc-AI/TeleOCR)

## Sources

- [StarDoc-AI/TeleOCR](https://huggingface.co/StarDoc-AI/TeleOCR) — Hugging Face, Aug 14, 2026

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "StarDoc-AI Releases TeleOCR for Document Parsing", The Open Weights, Aug 14, 2026, https://theopenweights.com/news/teleocr-huln