SoMark OXR parser: benchmarks and commercial license limits | Magica
SoMark releases OXR document parser with $2 million licensing thresholds
Editorial Team
••📖7 min read
The team behind document-parsing service SoMark has released OXR-1.0 code and weights. A benchmark retest supports a strong result, but speed and accuracy were measured separately, and users above funding or revenue thresholds need a commercial license.
OXR converts PDFs and document images into Markdown or JSON, including layout coordinates.
OmniDocBench maintainers retested it at 97.03, below the 97.18 score in SoMark’s promotional comparison.
Users exceeding $2 million in consolidated funding or annual revenue need a commercial license, subject to exemptions.
The team behind SoMark, a document-parsing service for AI agents, has released OXR-1.0, a model developers can run on their own infrastructure. OXR—Optical Everything Recognition—extracts document content into Markdown or JSON with layout coordinates. Access comes with a commercial boundary: its custom license requires a separate agreement for users above specified funding or revenue thresholds.
In its announcement, the team says its existing SoMark service uses a separate model, supports more file formats, recognizes document elements in finer detail and processes documents faster. It also offers additional post-processing, custom development and model fine-tuning. The downloadable model sits alongside that broader service offering and uses compatible Python and JavaScript clients.
For developers, the consequential choice is whether a downloadable parser can reliably handle their own documents at an acceptable operating cost. OXR offers replaceable layout analysis and a method for generating targeted post-training data. Its strongest speed and accuracy claims, however, come from different test configurations, and one benchmark has already exposed a tradeoff between the advertised score and the team’s preferred real-world settings.
SoMark’s company-reported parsing example shows a fictional résumé, layout annotations and a Markdown preview. Source: SoMark. Source: SoMark｜OXR.
What the benchmark retest changed
SoMark says OXR leads olmOCR-Bench, OmniDocBench and DocParsingBench. Its accuracy comparison is explicitly an October 5, 2026 snapshot combining published competitor scores with its own evaluations; not every model was rerun in the same environment.
OmniDocBench evaluates document parsing across text, formulas, tables and reading order. Its submission and retest discussion supplies a subsequent check. On October 19, OXR’s submitter reported an overall score of 97.18 on all 1,651 pages of OmniDocBench v1.6_full, using the official end-to-end evaluation process and quick_match matching method.
Internet scans and leaked databases expose the spread of ARTEX, a security-testing framework used in cyberattacks. Researchers documented largely autonomous compromises, but the records also show failed cashouts and uncertain counts of independent deployments.
Editorial Team
After a maintainer retested the model on an A100 machine, the submitter explained on October 28 that the layout-analysis confidence threshold had changed. The original evaluation used 0.25. The team subsequently raised it to 0.5 because the lower setting caused some duplicate recognition, and said the higher threshold worked better in real use.
Keeping other settings fixed, SoMark reproduced 97.18 at the lower threshold and 97.03 at the higher one. The maintainer’s A100 retest also produced 97.03 overall. SoMark attributed the reduction chiefly to differences between its formula segmentation and the benchmark’s annotations, and accepted the retested result. A maintainer said the leaderboard had been updated on October 29.
SoMark’s promotional page still displays 97.18 and warns that its comparison does not itself establish official leaderboard inclusion. The later issue documents both inclusion and the lower accepted score. That accepted result remains above rival parser TeleOCR’s 96.91 overall in SoMark’s frozen comparison, although the juxtaposition is not a fresh evaluation of both models under identical conditions.
Nor does the company’s comparison show a sweep of every component. It lists OXR’s table-structure-and-content TEDS score at 96.62, behind TeleOCR’s 96.82. TeleOCR also has a lower, better reading-order edit distance: 0.1184 against OXR’s 0.1214. A buyer’s mix of tables, formulas and text matters more than an aggregate ranking alone.
The company-reported evaluation was followed by a maintainer retest and confirmation of a leaderboard update. Milestones are shown in chronological order. Source: 请求收录 OXR-1.0 的 OmniDocBench 评测结果 #264.
A replaceable layout model, with precedents
The project documentation lists 13 supported document-element types, including tables, formulas, code blocks, chemical equations and chemical structures, as well as ordinary text and images.
Its pipeline uses a separate layout-analysis model to identify regions and their boundaries. OXR recognizes the cropped content and predicts reading order from the full page and region information; the outputs are then joined by region ID. Developers can replace the layout model while retaining OXR’s recognition and reading-order capabilities. This matters when a business needs different rules for grouping or classifying the same page.
SoMark calls the independent, pluggable layout module a first. The broader idea of separating layout analysis from content recognition predates OXR: a September 2025 technical report described MinerU2.5, a 1.2-billion-parameter document-parsing model, analyzing global layout on downsampled images before recognizing content on native-resolution crops. That is precedent for staged parsing, although it does not establish the same interchangeable-module interface OXR offers. SoMark’s announcement also credits Ovis2.5 as inspiration for OXR’s internal model design.
The other proposed distinction is recursive self-improvement, or RSI, during post-training. According to SoMark, a general vision-language model assesses OXR’s predictions, while disagreement among repeated OXR outputs helps identify uncertain samples. Explanations of failures are clustered, then a general model generates rendering and layout code to create new labeled document images for subsequent training.
This is a post-training data-production loop, not a claim that a deployed parser autonomously learns from every uploaded document. SoMark says new training examples require no manual labeling or correction, but validation and filtering remain part of the process. Unlabeled real documents expose weaknesses and supply style references; they do not automatically become trustworthy labels. The repository says a technical report will be added when published.
Downloadable weights do not settle deployment costs or rights
SoMark reports that OXR processed 100 fixed pages in about 101 seconds on average in a speed comparison of 21 models, with throughput 33.5% above the next-fastest model. The hardware was a single NVIDIA RTX PRO 6000 Blackwell with 96GB of memory; eight warm-up pages were excluded.
The company expressly says accuracy and speed were measured on different document sets with different runtime configurations. Equivalent accuracy under the speed configuration has not been confirmed. Those results therefore do not establish the cost of achieving the advertised accuracy in a production workload.
The deployment instructions specify 24GB of GPU memory for low-concurrency testing and recommend at least 48GB for longer documents or higher concurrency. They require Linux and a BF16-capable NVIDIA GPU. SoMark says the official vLLM inference software does not yet support OXR directly, and supplies an out-of-tree plugin that requires no recompilation. The model-hosting page labels the stored model “2B params” alongside the team’s 1.3-billion-parameter description; the release materials do not explain the difference. The stated hardware requirements provide a more direct starting point for deployment planning.
The license version available at the research cutoff identifies Shanghai Shoukou Technology Co., Ltd. as licensor. Its additional terms take precedence over Apache 2.0. A separate written commercial license is required when users and their affiliates, considered together, exceed either $2 million in total funding or $2 million in annual revenue—even for internal use or free services. Personal use and noncommercial research, including benchmarking, evaluation and teaching, are exempt from that threshold provision.
Funding includes equity and debt raised from any source; revenue means gross revenue for the most recently completed fiscal year. Both measures exclude intra-group transactions. Within those thresholds, the terms allow commercial activity, including paid products and services, without a separate commercial license.
Users providing products or services to third parties without a separate commercial license must display linked SoMark OXR attribution under the specified conditions. Failure to obtain a required commercial license or comply with attribution requirements automatically terminates rights under the license. The document does not quote a commercial-license price.
The next test is the customer’s document set
There is already an integration beyond SoMark’s tools. The oar-ocr project, which supports OCR and document parsing through multiple models, added OXR-1.0 alongside OvisOCR3 to its model-ID-based parser in its October 26 release. Earlier releases already supported alternatives including MinerU, PaddleOCR-VL and TeleOCR. This gives developers another implementation route and alternatives to compare; it does not demonstrate customer adoption or production reliability.
SoMark also offers an online demo and a free API without registration or an API key. It cautions that requests are queued, service stability is not guaranteed and production use is not recommended. Its Python and JavaScript clients can use either that API or a self-hosted deployment by changing the endpoint.
For a prospective deployment, the missing evidence is extraction quality, sustained throughput, memory use and operating cost measured together on the intended document set. Users above the license thresholds also need commercial terms before proceeding beyond exempt uses. The release makes evaluation accessible; choosing it for production still requires those workload results and, where applicable, an agreement with the licensor.
Meta's strategy chief and former CFO is joining Fireworks AI ahead of a reported fundraising push targeting at least a $30 billion valuation. The inference company announced a $17.5 billion valuation in July, making the hire part of a much larger expansion than its older funding figures suggest.
Sand.ai says its $200 million Series A+ round lifts its valuation above $1 billion. Co-lead investor Wangsu plans inference and deployment cooperation, while its filing shows substantial losses behind the startup’s video AI ambitions.
An advocacy group estimates Interior Secretary Doug Burgum’s AI holdings at $18.1 million–$64.4 million as his department advances data centers on public land. His ethics agreement allows retained investments with recusals, while pending permits will determine both the application of those safeguards and the public’s return. Interior denies his decisions are driven by personal investments.
Microsoft is evaluating fewer neocloud partners while expanding its own infrastructure. Delivery-dependent contracts underpin supplier financing, while IREN reports rising revenue alongside a $6 billion funding target for its broader expansion.
Physical Superintelligence says its AI system found a proof of a conjecture limiting quantum query advantages on random inputs. Public proof files enable scrutiny, but do not establish the private system’s cost or repeatability.
OpenAI is expanding its relationship with Hacktron, whose researchers previously reported an attack chain into employee accounts, to test third-party connections with greater access to cybersecurity models.
Google has acquired Crusoe’s southeast Wyoming data center site, where it plans a long-term $17 billion investment. Crusoe describes completed pre-development work; independent reporting points to customer setbacks. Power delivery and carbon capture remain key tests.
OpenAI is reportedly developing a Business Hub with daily mention and click charts, product updates and business agents. The unannounced feature could give verified businesses direct visibility into ChatGPT responses, without revealing prompts or competitor rankings.
Palantir and Nvidia have specified an integrated stack for running AI in customer-controlled environments. The architecture builds on existing software and infrastructure offerings, with published configurations spanning 32 to 640 GPUs, while pricing and workload economics remain unproven in the launch materials.
Democrats’ incoming congressional majorities put hearings and investigations within reach, but AI rules and crypto legislation still face bipartisan bargaining and a Republican president.
Microsoft reportedly declined to renew Oracle server rentals as older chips became less useful to it. The decision comes despite continuing capacity shortages, while Oracle says demand for earlier-generation GPUs remains strong.
An investigation puts Nvidia’s future commitments and guarantees at $530.5 billion. Its disclosures show rapidly expanding supplier commitments and phased support for customers whose infrastructure could generate further chip sales, with cash payments depending on timing, demand and contract terms.
TSMC is reportedly planning a possible move into 3D DRAM at Nvidia’s request. The foundry already helps integrate AI processors with memory, but a manufacturing expansion has no disclosed factory or production date.
Physicl has raised a $10 million seed round and introduced Studio, a beta app for generating targeted physical AI data. Its reported customer gains concern 3D reconstruction, leaving broader robot reliability and commercial economics to be demonstrated.
New cloud partners in Southeast Asia must enroll in Nvidia’s Fleet Intelligence system, according to a report, linking preferred chip access to monitoring while leaving data sharing and existing partners’ obligations unresolved.
Airbnb sold intellectual property to a private AI lab controlled by CEO Brian Chesky in a $7.1 million arrangement that also reimburses expenses. It obtained a license to certain lab technology and future investment rights, but disclosed no asset valuation or detailed license terms.
Chip shares recovered part of Thursday’s losses in early Friday trading. Fears about data-centre construction face a test against AI-chip supplier Nvidia’s strong reported sales, growth forecast and conditional infrastructure-financing plans.
SoftBank shares fell 10.8% as default-insurance costs reached a four-year high. Delayed listings and unfinished data centers complicate its AI investment outlook, while a $40 billion bridge facility has a March 2027 maturity and a repayment plan based on existing assets and other financing.
U.S. prosecutors allege a global fraud combined AI impersonation, sham confidential deals and cryptocurrency laundering. A blocked bank transfer shows why convincing an employee and moving the money are separate hurdles.
