Chinese OXR: an open-source 1.3B-parameter model that reads documents better than anyone
On October 29–30, the Chinese team SoMark released OXR into the open — a document-recognition model with just 1.3 billion parameters. It took first place in three separate global benchmarks for parsing PDFs and scans, beating systems many times larger, and it also came out fastest in the open-source class. The bigger story, though, is not the scores but how it was trained: nobody labeled the data by hand. The model was made to keep teaching itself in a loop.
What was released
OXR stands for Optical Everything Recognition. It takes a PDF or a photo of a document and returns clean Markdown or JSON, with the coordinates of every block on the page. It understands 13 element types: text, headings, tables, formulas, code, images, captions, footnotes, headers and footers, stamps, chemical equations and even chemical structures.
According to the developers, OXR is the first document-parsing model whose layout-analysis module is separate and swappable. That matters less for architectural elegance than for real work: different customers need different page-layout rules, and you can now drop in your own layout module while keeping OXR's content recognition.
Three first places, and the speed
The scores: 97.18 on OmniDocBench, 87.84 on olmOCR-Bench and 87.76 on DocParsingBench — first place in all three. For comparison, on OmniDocBench it edges past Gemini 3 Pro (92.91), GLM-OCR (95.22) and MinerU2.5-Pro (95.75), while being a fraction of their size.
Speed is the other half. On a single RTX PRO 6000, OXR parses 100 pages in 101 seconds — about a page per second, end to end, straight to finished structure. That is a first among open-source parsing models, at 557.77 native tokens per second by the team's measurements. More important: a 1.3B model fits on an everyday card such as an RTX 3060, so you can run it locally without renting a cluster.
How the model teaches itself: the RSI loop
Document-parsing accuracy is usually pushed up with manual labeling: people fix errors, the model is retrained, a new version ships, new messy scans appear, and the cycle repeats. The trouble is that perfect labeling of documents is impossible — layouts are endlessly varied and humans err as often as machines.
The team noticed something easy to miss in this class of task: document parsing almost always has an objectively correct answer. A line swapped, a table broken, a formula symbol lost — all of that can be checked exactly. So the human can be removed from the loop. That is RSI, recursive self-improvement, applied to document parsing — a first, the authors say.
It runs in three steps. First OXR parses a mass of unlabeled documents, and a helper multimodal system sorts the results into three buckets: correct, wrong, and uncertain. The third bucket is not a nicety: if you keep only right/wrong, the model's hesitation gets forced into a confident label and poisons the next training round. Each document is also run several times; if the outputs diverge, the sample goes to “uncertain” even when one run happened to be right. Then the disputed examples get mild image perturbations — scale, compression, slight blur — to see whether the error grows, which exposes the model's weak spots.
In step two, a diagnostic model describes the cause of each error in words: “cross-column cell alignment failure,” “handwritten cursive misread.” Those reasons are clustered to reveal what is missing from the data. In step three, a generator deliberately fills the gaps: it uses real layouts as the base, injects the problem structures, and re-tests the visual weaknesses it found. The new data goes back into training, and the loop repeats. No human writes the labels.
Why this matters right now
Document parsing was long seen as dull plumbing. In the age of AI agents, it has suddenly become a bottleneck. The most valuable professional knowledge — reports, scans, tables, manuals, medical forms — sits in PDFs and photos with chaotic layouts. Feed a model garbage with scrambled line order and collapsed tables, and you get confident nonsense, no matter how many parameters it has. So a tool that turns a messy document into clean structure is not “just OCR” — it is the front door to useful AI work.
And here we hit the same wall as everywhere else: input quality matters more than model size. Across anonymized data from our production database over the last 30 days, users started 14,925 image generations; 14,188 completed and 737 failed — about 95% success. But of those 737 failures, 541 (73%) were safety-filter decisions, not technical crashes. In other words, the model refuses over the content of the request more often than it fails to render. The same principle, mirrored: what goes in matters more than raw power at the output.
Speed, on our side, has also stopped being the problem: across 13,650 successful generations that month with a recorded completion time, the median was 72 seconds and the 90th percentile 187 seconds. The wait is short; the bottleneck shifts to what we feed in and how.
How to use it, and where we fit
OXR is fully open: the code, weights and the entire evaluation suite are on GitHub and Hugging Face, and the team has put up a free API — no sign-up, no key — with an honest note that it stays online as long as the servers hold. For local use you need a BF16-capable GPU, and it installs via the official Docker or CLI.
If you want to put document parsing to work right now, we have our own tools. You can upload a file and ask questions about it, tables and text included, in our chat — it accepts documents and reads their contents. And to transcribe a lecture, interview or meeting and get text with subtitles, there is the free audio/video-to-text tool. For images where text must come out clean, use models that handle lettering carefully, such as ChatGPT Images 2.5.
What it means in practice
The race for “who has more parameters” is slowly giving way to “who handles data more carefully.” A 1.3B model that beat far larger rivals not by size but by a smart retraining loop and a clean architecture is a good illustration. If open document-parsing tools keep getting cheaper and faster, good AI stops being a privilege of those who can afford the biggest cluster. That may matter more than one more line in a leaderboard.