AI News Feed
Market watch
New Mexico Clinic Trains Bilingual Health Apprentices to Address ShortageApple launches OLED iPad mini, but 60Hz display draws criticismAkamai Expands AI Cloud With Managed Inference and Agent SandboxAndroid Authority Highlights Oukitel WP63 Igniter and OPPO Find X10 Pro MaxAnt Group launches Lingguang 2.0 personal AI agentChangXin Rejects Apple Orders as Anthropic Files for Public Listing
New Mexico Clinic Trains Bilingual Health Apprentices to Address ShortageApple launches OLED iPad mini, but 60Hz display draws criticismAkamai Expands AI Cloud With Managed Inference and Agent SandboxAndroid Authority Highlights Oukitel WP63 Igniter and OPPO Find X10 Pro MaxAnt Group launches Lingguang 2.0 personal AI agentChangXin Rejects Apple Orders as Anthropic Files for Public Listing
Products & Applications

Chinese Teams Open-Source Long-Document OCR Models, Claiming Benchmark and Efficiency Gains

SoMark's 1.3B-parameter OXR claims first place on three document-parsing benchmarks and faster full-document parsing than PaddleOCR-VL-1.6 and MinerU2.5-Pro, QbitAI reported. Baidu's Wenxin Unlimited OCR 2 claims about 96% lower KV-cache use and continuous parsing of thousands of pages.

AI News Feed · 2026-10-27 20:42

OXR, short for Optical Everything Recognition, scored 97.18 on OmniDocBench, 87.84 on olmOCR-Bench and 87.76 on DocParsingBench, according to evaluation results released by the SoMark team. The team, which QbitAI described as coming from a Chinese startup focused on document intelligence, said OXR parses 13 types of document elements including text, tables, formulas, chemical structures and code blocks, outputs Markdown or JSON, preserves layout coordinates and restores reading order in multi-column pages. Tables are output as HTML; formulas as LaTeX; chemical equations as mhchem; chemical structures as SMILES.

In a same-condition full-parsing test on an RTX PRO 6000 Blackwell 96GB GPU, OXR took about 101 seconds on average to parse 100 pages randomly drawn from OmniDocBench, compared with about 145 seconds for PaddleOCR-VL-1.6 and about 255 seconds for MinerU2.5-Pro, QbitAI reported, citing the team's results. OXR uses a lightweight layout-analysis model to locate content regions and a main model to recognize content and determine reading order, a division of work the team said reduces the overhead of handling reading order separately.

The team said it introduced a Recursive Self-Improvement, or RSI, training approach in post-training for document parsing. OXR first parses unlabeled documents, external models evaluate the results and analyze errors, a separate clustering model groups the problems, and external models generate targeted training data for the next round. The added data does not require manual annotation, according to the team. OXR's layout analysis is also designed as a pluggable module, so developers can use the built-in model or connect another layout analyzer while reusing OXR's content recognition and reading-order functions.

Baidu's Wenxin Unlimited OCR 2, also released on Oct. 27, targets a different constraint: the memory cost of processing very long documents. QbitAI reported that the model can continuously recognize and transcribe the seven Harry Potter novels, more than 3,600 pages, and the Three Body trilogy, more than 1,200 pages. In a 100-page single-sequence test in its technical report, peak KV-cache use fell from 47GB under standard decoding to 1.77GB, a reduction of about 96%, with model weights and other memory costs excluded.

The model introduces a Sparse Reference Sliding Window Attention mechanism, or SR-SWA, with two filtering stages. A Page Gate first narrows the candidate visual information to pages relevant to the current parsing position, and a Token Sieve then selects a fixed number of relevant visual token blocks from those pages. A fixed-size local window carries recent context as generation proceeds. Baidu also uses an Unlimited-Horizon Decoding Paradigm that splits a document into page chunks, overlaps adjacent chunks by one page, reuses the overlapping page's visual KV cache and releases caches no longer needed.

QbitAI said it tested the model on an 884-page English edition of The Lord of the Rings; parsing took about 42 minutes 56 seconds, or about 2.9 seconds per page. The output had 884 pages, text on pages 600 and 750 matched the original, and the system marked 466 cross-page merges, including a sentence split between pages 873 and 874 that was rejoined. In the technical report, the model scored 84.59 on MPDocBench and 77.06 on MarathonDocBench, a new 100-document benchmark with 103 to 300 pages per document; the report said both were the highest composite scores among the models listed.

Baidu reported that the chunked decoding approach raised end-to-end throughput in a 16-way concurrency test from 360 to 1,315 tokens per second, about 3.65 times standard decoding. SR-SWA's single-step decoding speed reached 7.0 times that of full attention and 5.6 times that of R-SWA used in the previous generation, according to the report. The previous Unlimited OCR had more than 10 million downloads and about 28,600 GitHub stars, QbitAI reported. Both OXR and Wenxin Unlimited OCR 2 have opened their code and model weights, with Baidu's model available on Hugging Face and GitHub and OXR offering an online demo and free API.