OXR: 1.3B Open-Source Parser Tops Three Doc Benchmarks at 3.05x GLM-OCR Throughput
rohanpaul_ai · x · 2026-10-21
OXR is a 1.3B-parameter open-source document parser, claimed as the first to use RSI (self-improvement) in post-training: a VLM evaluates its own outputs, rollout disagreements flag uncertainty, and a separate clustering model groups failure causes to guide targeted data generation for the next round. Results: #1 on olmOCR-Bench, OmniDocBench and DocParsingBench, with 3.05x GLM-OCR's throughput on a single RTX PRO 6000 — the fastest open-source result in the speed comparison. Positioning: turn documents into agent-friendly data. Code is on GitHub.
More from Multimodal
- Flux Klein's reference latent node excels at style transfer — can Qwen 2.1 do it? — Inevitable_Pen9043 · 2026-10-21
- AI-generated character IP demo: one mascot across posters, merch and conference scenes — eyal_zuri · 2026-10-21
- Cinematic AI fantasy romance short 'Alpha Meik & Luna Martina' released — Soundangel-Martina · 2026-10-21
- AcademiaSD adds Fibo 2.0 LoRA training with NF4 quantization, cutting VRAM from 16GB to 5.2GB — AcademiaSD · 2026-10-21
- "We went deeper and found more sky": stunning AI-generated megastructure sci-fi art — piotrbinkowski · 2026-10-21
- AI reimagines the 1993 skating game Skitchin with strikingly faithful 3D visuals — demondisc · 2026-10-21