TeleOCR
by StarDoc-AI
A 1.2B document parser for both clean PDFs and phone photos of paper
StarDoc-AI/TeleOCRDeploy TeleOCR
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
TeleOCR is an open-source vision-language model that turns a page image into structured text: paragraphs, tables and formulas, in reading order. Most document parsers are tuned for clean digital pages or for photos of paper. TeleOCR is trained for both in one model, including pages that are curved, skewed or photographed at an angle. StarDoc-AI released it in August 2026 as NaviDC-OCR and renamed it TeleOCR on 10 September 2026. It is Apache-2.0 licensed.
On OmniDocBench v1.6 the card reports 96.87 overall, the highest in its comparison table, with 97.05 table TEDS and a 0.027 text edit distance. The same table lists Gemini 3 Pro at 92.85 and GPT-5.2 at 86.52. On PureDocBench it reports 86.90 on clean pages.
These are self-reported results on public benchmarks, and the repository showed about 32,000 downloads in its first month. Parsing quality varies with layout, so test it on your own documents before relying on the ranking.
Architecture
A lightweight vision-language model specialised for document parsing: about 1.2B parameters per the card, 1.42B in the safetensors index. The card lists geometry-aware modelling for camera-captured pages, content-structure decoupled learning for tables and formulas, pseudo-labels from multi-node consensus voting, image-to-image self-verification for data refinement, and a four-stage progressive training pipeline.
Mixpeek SDK Integration
# No Mixpeek extractor runs these weights. Run TeleOCR with the Quick Start on its
# model card to turn each page image into text, then index that text: store it as a
# text object in a bucket, and a collection's text extractor embeds and indexes it.
import requests
requests.post(
"https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
json={
"key_prefix": "filings/2026-q2/page-14",
"blobs": [{"property": "body", "type": "text", "data": page_text}],
},
)Capabilities
- Parses digital and camera-captured document pages with one model
- Outputs text, tables and formulas with reading order
- Handles curved, skewed and angled photos of paper
- About 1.2B parameters, Apache-2.0 license
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| OmniDocBench v1.6 | Overall | 96.87 | Model card: StarDoc-AI/TeleOCR (self-reported) |
| OmniDocBench v1.6 | Table TEDS | 97.05 | Model card (self-reported) |
| OmniDocBench v1.6 | Text edit distance (lower is better) | 0.027 | Model card (self-reported) |
| PureDocBench (clean) | Overall | 86.90 | Model card (self-reported) |
Performance
We have not measured throughput or memory. At about 1.2B parameters it is small enough for a single GPU; the card also points to a community GGUF build for llama.cpp.
Common Pipeline Companions
Frequently Asked Questions
What is document parsing, and how is it different from OCR?
OCR reads the characters on a page. Document parsing also recovers the structure: which text is a heading, which cells form a table, where a formula starts, and the order a person would read it in. Search and RAG need that structure, because a table flattened into lines of text loses which number belongs to which row.
Can TeleOCR read photos of paper documents?
Yes, that is its main difference from most small parsers. The card describes geometry-aware modelling for camera-captured pages and shows results on curved and skewed pages from the DocUNet and DIR300 datasets.
How does TeleOCR compare with Gemini or GPT on document parsing?
On the card's OmniDocBench v1.6 table it scores 96.87 overall against 92.85 for Gemini 3 Pro and 86.52 for GPT-5.2. These are self-reported numbers on a public benchmark, so they can be reproduced, and results on your own layouts may differ.
Does Mixpeek run TeleOCR?
Not on the managed tier. You can run it yourself and index its text output in Mixpeek, or on a single-tenant Enterprise deployment upload the weights and run them as a custom plugin.
Specification
Research Paper
TeleOCR technical report (arXiv 2608.12898)
arxiv.orgBuild a pipeline with TeleOCR
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free