tensorfolk

Baidu open-sources Unlimited OCR 2, a 3B model that parses hundred-page documents in one request

Baidu released Unlimited OCR 2 under the MIT license, a roughly 3B-parameter model that reads very long documents in one pass. A sparse sliding-window attention scheme keeps only a small fixed set of visual references active and drops stale cache, so memory stays bounded as pages pile up. Baidu claims top scores on multi-page parsing benchmarks, not yet independently checked.

Why it matters

A small, permissively licensed model for whole-PDF extraction is a cheap building block for document pipelines and RAG ingestion.

Source: Read the full story on MarkTechPost

ocrbaiduopen-sourcelong-documentsmit-license

More in Computer Vision

Computer Vision / Vision-Language Models

Tencent's WeVisOne-VL handles detection, OCR, segmentation and matting in one model

WeChat Vision researchers describe a vision-language model that covers detection, grounding, counting, OCR, segmentation and alpha matting by predicting structured tokens, with a small pixel decoder for masks. They claim top results across more than 30 benchmarks and say weights are being released.

Why it matters: One small open model could replace several separate perception models in vision pipelines.

Read on arXiv
Computer Vision / 3D & Spatial AI

Rivian buys Third Dimension AI, maker of the SuperSim Gaussian-splat driving simulator

Rivian has closed a deal for Third Dimension AI, whose SuperSim tool rebuilds real driving scenes from camera, LiDAR and IMU logs using Gaussian splatting and world modeling, then renders new viewpoints and inserted objects. The roughly 15-person team joins Rivian's autonomy group; terms were not disclosed.

Why it matters: Another automaker is betting on reconstructing real-world sensor logs as 3D sims, rather than hand-built scenes, to train and test autonomy stacks.

Read on Radiance Fields