Baidu open-sources Unlimited OCR 2 document model

Baidu has released Unlimited OCR 2, a 3.34-billion-parameter vision-language model for optical character recognition, as open source under the MIT license, according to aitimes.com. The model uses a sparse attention technique called SR-SWA that Baidu says keeps compute and memory use from growing sharply with long documents.

Baidu reports that Unlimited OCR 2 scored 77.1 on the MarathonDoc benchmark for documents running to hundreds of pages, against 33.7 for the comparison model Kimi K3.

Separately, Ant Group's InclusionAI released Ling-3.1-flash-1019 as open source under the MIT license. The model has 560 billion total parameters and activates 25 billion per token through a mixture-of-experts design, with support for up to 1 million tokens of context. Its AutomationBench-AA score rose from 62% to 68%, its AA-Briefcase Elo from about 1400 to 1524, and its Artificial Analysis Intelligence Index from 41 to 44.

The vLLM Semantic Router project released d3-ultra, an approximately 26.1-billion-parameter multimodal decision model, under the Apache 2.0 license. It takes text, images and video and picks one option from a set of choices, answers yes/no, or outputs probabilities, and can localize objects in images and track them in video. The team said answering 1 to 255 questions about a single image took 0.3 to 2.16 seconds.

US medical AI startup Concurrence introduced the Health World Model, which combines electronic health records with information gathered by AI agents talking to patients to build per-patient digital twins. The company is working with Databricks to build systems based on individual medical institutions' data.

Nace AI launched Drex for Search, a retrieval solution that uses a decision model to judge which folders, files, pages and tables likely hold an answer. It scored 92.05% accuracy on the BrowseComp Plus benchmark with an average of 12.6 search calls per question, placing second on that benchmark. In Nace AI's own test of 529 work-document questions, attaching Drex to existing coding agents cut task steps by 62% and input tokens by 56% and made processing 2.3 times faster.

Cognition launched Devin Security 2.0, which uses frontier AI models including Anthropic's Claude Mythos to find and fix software vulnerabilities. The product analyzes source code, prioritizes vulnerabilities by exploitability and impact, reproduces and patches them in a separate virtual environment, runs tests and opens a pull request, using an Agentic MapReduce structure.