Deep Dive — OXR beats MinerU on OCR, but its licence ends at $2M

A 1.3-billion-parameter document parser called OXR went to the top of three published benchmarks within three weeks of its repository being created, and you can download the weights. The scores are the vendor's own. The licence is narrower than the word "open" implies — and for the developers most likely to adopt a parsing model, the licence is the part that decides the question.

A three-week-old repository at the top of a crowded table

The launch was reported in Chinese on October 27, in an article QbitAI attributes to the SoMark team — "provided by the SoMark team, republished by QbitAI with authorisation." Read the prose as the company's account of its own model. The artefacts, though, are public: the SoMarkAI/OXR repository was created on October 5, and by this week it carried 1,461 stars and 95 forks. That is a fast start and a small one beside the tools it is aimed at: PaddleOCR's repository carries 91,759 stars, MinerU's 82,581.

Document parsing has been the least glamorous bottleneck in enterprise AI, and it has been through two architectures in eighteen months. The first was a pipeline: detect text regions, classify each block as text or table or figure, run a recogniser over every crop, then stitch the results back into reading order with hand-written rules. It is fast and brittle, and it fails in the way that costs the most — a merged table cell shifting one column left, so the revenue figure lands under the wrong year.

The second architecture, now dominant, hands the whole page image to a vision-language model and asks for Markdown. OXR's own comparison table doubles as a record of that transition, and the pattern in it is consistent: every generation buys one to three points by moving more of the layout job inside the model. On OmniDocBench, PaddleOCR-VL scored 94.18, then 94.93, then 96.34 for version 1.6. MinerU-2.5 sits at 93.04 and MinerU2.5-Pro at 95.75. On olmOCR-bench, DeepSeek-OCR opened at 75.70 and DeepSeek-OCR-2 reached 90.25.

The number in that table worth pausing on is not the leader's. It is Gemini 3 Pro's 92.91 and Gemini 3 Flash's 92.62 on OmniDocBench — roughly four points behind a 1.3-billion-parameter specialist. On a page of multi-column text with footnotes and a formula block, a frontier general model reads the words and loses the structure. That gap is the market OXR is selling into: not "can a model see text," but "can it recover the shape of a document faithfully enough for an agent to act on."

What OXR actually is

OXR stands for Optical Everything Recognition, and the name is the design brief: thirteen element types, not just characters. Text, titles, captions, footnotes, page headers and footers, tables, figures, formulas, code blocks, stamps, chemical equations and chemical structures. Tables come out as HTML or Markdown, formulas as LaTeX, chemical reactions in mhchem syntax and chemical structures as SMILES strings, so a reaction scheme on page six becomes something a downstream script can parse rather than a picture of a reaction scheme.

The pipeline is two models, and the split is the interesting engineering decision. A lightweight layout model finds the regions; OXR itself does content recognition and reading-order judgement in a single pass. Most rivals either fold layout into the recognition model or run a separate model that also owns reading order. SoMark's argument for the third arrangement is practical: layout conventions vary more by industry than any other part of the job — an archive wants a group of scans kept whole, an image-search product wants each one cropped — so layout is where customers ask for changes. Making it a pluggable module means a team can swap in its own layout model and keep OXR's recognition and ordering.

The training method is the part the announcement leans on hardest. After the base model, SoMark ran a recursive self-improvement loop: OXR parses unlabelled documents, an external model grades the output and diagnoses per-sample failures, a separate clustering model groups those failures by cause, and an external model generates training data aimed at the clusters. No human annotation sits in the loop.

That is a real technique and a well-established shape — error-targeted synthetic data, iterated — but "recursive self-improvement" is doing more marketing work than technical work here. Nothing in the described loop has the model rewriting its own weights or setting its own objectives; a critic model and a data generator stand outside it. The honest description is automated curriculum building, and the honest claim is that it is cheaper than paying annotators.

The model card lists 1.53 billion parameters for the safetensors repository against the 1.3 billion the team quotes for the model. Neither document explains the gap; with a two-model pipeline, the difference is consistent with the layout module being counted in the repository but not in the headline figure.

Three leaderboards, ordered by who ran them

The claim is first place on olmOCR-bench, OmniDocBench and DocParsingBench, at 87.84, 97.18 and 87.76 respectively. Those are the numbers, in the repository's benchmark document. The independence behind them is not equal, and the difference matters more than the decimals.

olmOCR-bench is the strongest result and the weakest evidence. The 87.84 does sit first on Hugging Face's olmOCR-bench leaderboard, above Infinity-Parser2-Pro at 87.6 and LightOnOCR-3-4B at 86.3. But the leaderboard entry is submitted from the model owner's own evaluation file, not run by the benchmark's maintainers, and second place is 0.24 points behind across 1,403 pages and 8,413 scoring rules. That is a tie at the top, not a rout, and the write-up's framing of a model "sweeping" three charts is doing work the margin cannot support.

OmniDocBench is a genuine third-party protocol, self-run. The 97.18 is on 1,651 pages using OpenDataLab's evaluation, with peer scores taken from the OmniDocBench repository at a pinned commit — a cleaner comparison than most vendors publish. It is still SoMark running SoMark. A useful tell of how protocol-sensitive these scores are: our September brief on FLUX 3 Action and Sarvam's document-AI record put PaddleOCR-VL 1.6 at 96.01 on the v1.6 release of the same benchmark; OXR's table puts it at 96.34. A third of a point of daylight between two vendors' readings of one competitor.

DocParsingBench is SoMark's own. The 87.76 is first among 17 systems — on a benchmark whose toolkit and dataset live in a repository under the same SoMarkAI organisation, with the dataset released in March 2026, before OXR existed. That does not make it worthless: the data and scoring code are open, and a benchmark built from finance, legal, scientific, manufacturing and education documents is closer to a customer's actual files than an academic PDF set. It does make "first on three leaderboards" a weaker sentence than it reads, because one of the three juries works for the defendant.

Where the top ranking holds, and where it doesn't

The component scores are more informative than the totals, and they cut against the announcement in two places.

OXR's best result is text fidelity: an edit distance of 0.0251 on OmniDocBench, ahead of OvisOCR2 at 0.0265 and TeleOCR at 0.0267, and the cleanest measure of "did it transcribe the page correctly." Formula recognition at 97.44 CDM is second, behind Xiaomi-OCR-0 at 98.50.

Tables are where the headline claim is thinnest. The repository says OXR leads "with a particularly wide lead in table parsing." On OmniDocBench's table metric, TeleOCR scores 96.82 against OXR's 96.62 — OXR is second. On olmOCR-bench's table category it wins by 0.95 points over Chandra-OCR-2. The wide lead exists on DocParsingBench, where OXR's 80.04 beats Chandra's 77.69 — SoMark's benchmark, SoMark's table corpus, SoMark's winner.

Reading order is the sharper contradiction, because reading order is what the architecture pitch is about. OXR's 0.1214 edit distance on OmniDocBench is fourth: behind OvisOCR2 at 0.1120, Youtu-Parsing at 0.1160 and TeleOCR at 0.1184. The design decision the team presents as its central idea — folding ordering into recognition so the model judges sequence from meaning — is not where the model wins. It wins on transcription and on speed. Both are worth having; neither is the story the launch tells.

The speed result is the one to keep

Speed is where OXR's advantage is largest and least ambiguous. On a single RTX PRO 6000 Blackwell with 96 GB, parsing a frozen 100-page slice of OmniDocBench end to end, OXR finished in 100.97 seconds. OvisOCR2 took 134.8, PaddleOCR-VL-1.6 took 144.7 and MinerU2.5-Pro took 254.7 — about 0.99 pages per second against 0.39.

Put that at archive scale and it stops being a benchmark. A million pages at OXR's measured rate is roughly 11.7 days of continuous work; at MinerU2.5-Pro's rate it is roughly 29.5 days. Serial decoding runs at 557.77 tokens per second against PaddleOCR-VL's 516.39, an 8 percent edge that matters less than the pipeline result.

Two caveats belong with the numbers. The harness is SoMark's — one GPU, one vLLM version for its own model, batch settings not audited across 21 vendors — so treat it as an indicative ranking rather than a controlled bake-off. And the announcement compares against the two most-searched names, PaddleOCR and MinerU, while the actual runner-up in SoMark's own table is OvisOCR2, a model most readers have never heard of. Both framings are accurate. Only one is useful.

The licence is the part nobody translated

Here is what the English-language coverage has not carried. OXR is not under a standard open-source licence. It is Apache 2.0 plus additional terms, and the additional terms do three things.

Commercial use is free only while you and your affiliates, on a consolidated basis, stay under $2 million in total funding and $2 million in annual revenue. Cross either line — cumulative equity and debt, not revenue alone — and you need a paid commercial licence from SoMark before you continue using it. Run an online service on OXR and you must prominently credit SoMark OXR in the interface or public documentation. And you may not use it, at any size, to offer a hosted API that competes with SoMark's own document-parsing API — including through a reseller, a white-label arrangement, or an intermediary interface, paid or free. Breach any of it and the licence terminates automatically, without notice.

Internal use inside one organisation is explicitly carved out of the hosted-API ban. So a bank parsing its own loan files is fine, and a company under the thresholds building a RAG pipeline is fine. The businesses the ban catches are the ones most likely to want a fast, accurate parser: the seed-stage startup selling document extraction as an API, the vertical product whose parsing layer is the product, the agency wrapping it for clients. GitHub's own licence field reads "Other," which is the correct signal — the OSI definition does not permit field-of-use restrictions, and a funding threshold is a restriction on who may use the software at all.

Set that beside the incumbents. PaddleOCR ships under straight Apache 2.0, and so does datalab's Chandra, per both projects' repositories. MinerU carries its own custom terms as well, so SoMark is not alone in hedging — but it is asking for commercial licences from a class of user that its competitors do not. If you are picking a parser this quarter, that is the line item that outranks 0.24 points on a leaderboard.

The strategic read is straightforward once the licence is in view. OXR is not the product; SoMark's hosted parsing engine is. The weights are the distribution channel — a free model with a free API, which the team says it will keep running "until we can no longer afford to run it," pulls developers into an SDK that points at the paid service, and the licence makes sure nobody builds a cheaper API on top of it. It is the same template Meta used with Llama and that several Chinese labs now use: publish the weights, keep the pipeline.

What this means for the parsing market

The technical result is real and it points one way: small specialised parsers are beating general frontier models on document structure, and the margin is now large enough that "just use Gemini" is no longer a serious answer for anyone processing documents at volume. The competitive set has compressed accordingly — MinerU, PaddleOCR-VL, dots, Chandra, LightOnOCR, jina-ocr, Infinity-Parser and OXR are all within a few points of each other on OmniDocBench, and the differentiator is shifting from accuracy to pages per second, licence terms, and whether the layout module bends to your documents.

That compression is why the licence is the story rather than a footnote. When six models parse a page equally well, the terms become the product decision, and a parser is the front door to every document an enterprise owns — including the ones it would rather not hand to a vendor's API. Parsing is also the step that decides whether retrieval works at all; AI 101 — What is RAG? covers the plumbing downstream, and it is only as good as the pages fed into it. OXR's terms say, in effect: run it inside your walls, tell your users it is ours, and do not sell what we sell.

One absence is worth naming. The LlamaIndex and Kaggle ParseBench leaderboard, built on about 2,000 human-verified enterprise pages across five capability dimensions including charts and content faithfulness, listed 34 models when it was updated on October 20. OXR is not among them, and neither OmniDocBench nor olmOCR-bench tests chart-data extraction at all. The one independent benchmark closest to a customer's real workload has not seen this model yet.

What to watch

The technical report. The repository still carries an arXiv badge reading "coming soon" and asks readers to cite the repository until the paper lands, so "pretrained from scratch" and the RSI loop are claims without a methods section. Watch what training data the paper discloses and whether the RSI rounds are reproducible.

Independent reproduction of the speed run. The frozen 100-page list is published, which makes this unusually easy to check. Anyone with a 96 GB Blackwell card can settle the throughput question in an afternoon, and throughput is the claim OXR needs most.

Whether OXR turns up on ParseBench, and whether a second vendor's harness confirms the 87.84. A leaderboard position submitted by the model's owner is a claim; a run by the benchmark's maintainers is a result.

And the licence in practice. The interesting test is not whether a bank adopts OXR — it will, freely — but whether a seed-stage document-AI startup with $2 million in the bank builds a product on a parser it may have to license away from later. If the industry's next wave of parsing companies chooses Apache-2.0 PaddleOCR or Chandra instead, the funding threshold will have cost SoMark the thing it most needs: a developer base that cannot easily leave.

If you are choosing a document parser for production this year, would a benchmark lead or a clean licence decide it for you? Write to us via our contact page.