TeleOCR for Chemical Companies: A Free Model That Tops One Document Test
TeleOCR is a small, free model that reads documents. On OmniDocBench, the independent document test, it is level with OXR-1.0, the leader, on whole documents and on tables. An outside tester challenged TeleOCR's table scores on 26 October, and OmniDocBench has not ruled on it yet. A second test ranks it 98th. For a chemical distributor or producer reading supplier documents, the gap between the two says what to check before you rely on it.
Latest update
ParseBench added Reducto's r-1-mini above TeleOCR on everything but charts, so TeleOCR is now 98th of 156 overall instead of 97th of 155, 115th on tables instead of 114th, 9th instead of 8th on tracing text back to its place on the page, 110th on content instead of 109th, and 114th on formatting instead of 113th.
TeleOCR at a glance
- Made by
- XingChen-AGI, published on Hugging Face4
- Released
- 17 Aug 2026 as NaviDC-OCR; renamed 10 Sep 20264
- Size
- About 1.2 billion parameters4
- Price
- Free to download and run, Apache 2.0 license4
- Languages
- Chinese and English4
- OmniDocBench
- 1st of 37 on tables; 2nd of 37 on whole documents, level with OXR-1.01
- ParseBench
- 98th of 156 overall2
What TeleOCR is.
TeleOCR is a document reading model of about 1.2 billion parameters, small enough to run on your own servers. You give it a page as an image, and it returns the text, the tables, and the formulas in order. Its weights are free to download under the Apache 2.0 license.
Its makers released it on 17 August 2026 as NaviDC-OCR and renamed it TeleOCR on 10 September. They built it for both digital documents and pages photographed with a camera. Its model card lists Chinese and English.
How TeleOCR reads on OmniDocBench.
OmniDocBench scores whole pages: text, tables, and formulas together. It is run by OpenDataLab, which also makes MinerU, a free model it ranks; MinerU places fifth.
Whole-document reading score: text, tables, and formulas, out of 100
OXR-1.0 joined the test on 29 October, 0.12 points above TeleOCR. The test publishes no margin of error, so the two are level. OXR-1.0's license allows free business use only while a company has no more than $2 million in funding and $2 million in yearly revenue, so TeleOCR is the best result most chemical companies can run free. Gemini 3 Pro is 4 points back.
OmniDocBench leaderboard29 Oct 2026
At the top of one test, 98th on another.
ParseBench, run by LlamaIndex, which sells a rival tool, scores five things a document reader must get right. TeleOCR is near the top on one of them and in the bottom half on the other four. Each track runs from first place on the left to last on the right. ParseBench re-scores published results when its scoring method changes, so ranks can move.
OmniDocBench37 results
Whole documents
Text, tables, and formulas together; level with OXR-1.0, the leader
2ndof 37score 96.91Tables
Table structure; level with OXR-1.0
1stof 37score 96.82
OmniDocBench leaderboard29 Oct 2026
ParseBench156 results
Overall
The average of the five below
98thof 156score 57.52Tables
Merged cells and headers that span rows
115thof 156score 72.77Charts
Data points read off a chart
115thof 156score 9.97Content
Nothing left out, nothing made up, read in order
110thof 156score 84.59Formatting
Bold, strikethrough, superscripts, and headings kept
114thof 156score 45.74Tracing to the Page
Each piece of text traced back to its place on the page
9thof 156score 74.55
ParseBench leaderboard6 Nov 2026
TeleOCR for chemical distributors and producers.
Where its results point. No public test used a chemical company's own documents, so treat each of these as a job to test, not a result.
Read Supplier Documents
Pull the fields you need from a technical data sheet, a safety data sheet, or a certificate of analysis into your own systems.
Run it on 20 of your own documents, including a scan and a multi-page safety data sheet, and check every value against the page.
Pull Price Lists Into Rows
Turn a supplier's multi-page price list into rows you can quote from.
Check merged cells and footnotes. The second test scored it low on merged cells and headers that span rows.
Keep Documents In House
Read documents on your own servers when a file should not leave your company.
Someone has to own the setup. Compare its running cost with a paid reader's price per page.
Where TeleOCR goes wrong.
The second test finds gaps the first one does not score.
Charts Are a Blind Spot
It scored 9.97 out of 100 on reading data points from charts, 115th of 156. Do not let it read a chart on a spec sheet or a report unchecked.
ParseBench leaderboard6 Nov 2026
Formatting Gets Lost
It scored 45.74 on keeping bold, strikethrough, superscripts, and headings, 114th of 156. A superscript in a unit or a struck-out price can change the meaning of a line.
ParseBench leaderboard6 Nov 2026
Two Languages Listed
Its model card lists Chinese and English. A supplier document in German or Spanish needs its own test.
TeleOCR model cardNo date stated
You Run It Yourself
It is free, but it runs on your own servers, so someone has to set it up and keep it running. There is no published price to compare with a paid service.
TeleOCR model cardNo date stated
Questions about TeleOCR.
Short answers, from the same sources as the rest of this page.
What is TeleOCR?
A free document reading model of about 1.2 billion parameters, published on Hugging Face under the Apache 2.0 license. It was released as NaviDC-OCR on 17 August 2026 and renamed TeleOCR on 10 September.
Is TeleOCR the best OCR model?
On OmniDocBench, the independent document test, it is level with OXR-1.0 on whole documents and on tables, and OXR-1.0 needs a separate license at most chemical companies. An outside tester challenged TeleOCR's table scores on 26 October, and OmniDocBench has not ruled on it yet. On ParseBench, a test run by a company that sells a rival tool, it ranks 98th of 156 overall.
Is TeleOCR free?
Its weights are free under the Apache 2.0 license. You pay for the servers that run it.
Can TeleOCR read safety data sheets and certificates of analysis?
No public test used them. It reads whole documents and tables well on OmniDocBench, but scored low on charts and formatting on ParseBench. Test it on your own safety data sheets and certificates of analysis, and check every value.
Why do the two document tests disagree?
They use different documents and score different things. OmniDocBench scores text, tables, and formulas. ParseBench also scores charts, formatting, and tracing text to the page, and it ranks a product its maker sells.
Sources.
- OmniDocBench leaderboard OpenDataLab (Shanghai AI Lab), which also makes MinerU, a free model it ranks.Published 29 Oct 2026 · Vailent checked 2 Nov 2026
- ParseBench leaderboard LlamaIndex, which sells LlamaParse, a product it ranks.Published 6 Nov 2026 · Vailent checked 9 Nov 2026
- SoMark: OXR-1.0 model card and license The maker's own page. Its license allows free business use only while a company has no more than $2 million in funding and $2 million in yearly revenue.States no date · Vailent checked 2 Nov 2026
- TeleOCR model card The makers' own page.States no date · Vailent checked 30 Sep 2026