DeepSeek-OCR 3, served as an API
The next generation of the OCR model that reads pages the way an LLM reads text: send a page image, describe the output you want, get structured results in one step. Point the OpenAI SDK at our base URL and call it per token.
- Optical context compression
- Markdown or JSON out
- Priced per token, not per page
- Batch at 50% off
jk. It is not out yet.
DeepSeek has not announced a third-generation OCR model, and unlike the general DeepSeek line there is not even a rumor mill to report on. DeepSeek-OCR 2 is the current release and it is live here now.
Below is every rumor we could source, what DeepSeek has actually confirmed, and the model you can call in the meantime. When the real one ships, this page becomes its page at this URL.
The DeepSeek-OCR 3 rumor mill.
Each row labelled by how much it is worth. Last checked September 24, 2026.
- Release status
- ConfirmedUnreleased. DeepSeek-OCR 2 (“Visual Causal Flow”) is the current generation, and it cut repetition rates against the original from 6.25% to 4.17% on user-log images and 3.69% to 2.88% on PDF data. arXiv 2601.20552
- Leaks
- No signalNothing credible names a DeepSeek-OCR 3. No leak, no paper, no model card.
- What the line is built on
- ConfirmedThe original DeepSeek-OCR (“Contexts Optical Compression”) pairs a DeepEncoder with a 3B MoE decoder and reported 97% decoding precision on the Fox benchmark while text tokens stay within 10x the vision tokens. arXiv 2510.18234
- The weakness a v3 would target
- ReportedIndependent evaluation finds end-to-end optical compression trades away the visual robustness of a classic pipeline: on random non-semantic text it struggles where PaddleOCR-v5 holds precision. That is the obvious thing for a next version to fix. arXiv 2601.03714
- The name
- UnverifiedOurs. DeepSeek could fold OCR into a general multimodal release rather than ship a numbered successor, in which case this page follows that model instead.
- How it would serve here
- ConfirmedSelf-hosted on our own GPUs behind chat completions with image parts, the same shape DeepSeek-OCR 2 uses today, with batch at half rate.
- Price
- No signalUnknown until we have the model. Per token rather than per page, as with everything else in the catalog.
What you can call right now.
The DeepSeek models live on the API today, at the rates they bill at.
Full inference catalogDeepSeek-OCR 2
The current OCR model here. Send a page image, ask for markdown or JSON fields, get structured output in one call.
Input / output per 1M tokens
$0.039 / $0.039
DeepSeek V4-Flash
The same lab's general model, for work that needs reasoning over what the OCR pass pulled out.
Input / output per 1M tokens
$0.22 / $0.66
Why it shows up here early.
Not a promise about a date. A description of the path a new open model takes to get behind this API.
The OCR path here is already built
Page images in, structured markdown or JSON out, batch at half rate for archives. A new OCR model drops into a surface that exists rather than needing one.
Open weights make this a days problem
The gap between a Hugging Face release and a served endpoint is engine support and capacity, both of which we hold standing for this family.
Nothing to migrate when it arrives
Same base URL, same OpenAI SDK, same API key. Moving a document pipeline to the next OCR model means changing the model string.
One email, the day it is callable.
Leave an address and we send exactly one message when DeepSeek-OCR 3 is serving here, with the model id and the rate. Nothing else goes to this list.
DeepSeek-OCR 3, answered.
Is DeepSeek-OCR 3 out?
No. As of September 2026 DeepSeek has not announced a third-generation OCR model. DeepSeek-OCR 2 is current and it runs on OpenRelay today.
Are there any DeepSeek-OCR 3 rumors?
None worth repeating. Unlike the general DeepSeek line, which has an active rumor mill, nothing credible names a third OCR generation. The rows above are what the current models actually are, which is more useful than speculation.
When is the DeepSeek-OCR 3 release date?
Unannounced. DeepSeek publishes weights and papers with little warning; when a successor lands this page becomes the live model page and the notify list gets an email.
Will OpenRelay serve the next DeepSeek OCR model?
Assuming open weights, yes. We self-host the OCR model on our own GPUs, so a new release is engine work on our side rather than a vendor negotiation.
What should I use until then?
DeepSeek-OCR 2, live on the card above. It takes page images through the standard chat completions call and returns markdown or JSON, and the Batch API runs the same requests at half rate within 24 hours for archive-scale jobs.
Why does this page exist if the model does not?
Because people search for the next model before it ships, and landing on nothing is worse than landing on a sourced answer. This page says plainly that no rumors exist, tells you what the current models do, and becomes the real model page at launch.
More models
The account outlasts the model.
One API key, one base URL, per-token billing. Whatever DeepSeek ships next, adopting it is a one-line change rather than a new vendor.