OCRia-1: 4B doc-parsing VLM hits 0.87 mean ParseBench, decodes 2-4x faster
NielsRogge · x · 2026-10-26
LlamaIndex's team released a 37-page technical report for OCRia-1, a 4B vision-language model for document parsing, now on Papers with Code.
- Benchmarks: Averaging 0.87 on ParseBench across tables, charts, and text formatting vs 0.75 for the next-best system, beating a 35B model and Mistral OCR 4.1; SOTA on ParseBench-Tab.
- Capabilities: Handles merged cells, spanning headers, tables with thousands of cells, chart data extraction, and formatting fidelity (bold, italics, sub/superscripts).
- Speed: 4,700 tokens/sec decoding on one B200 GPU — 2-4x+ faster than latest open-weight models.
- Method: Performance attributed to careful data curation (synthetic pages for exact ground truth, real pages transcribed via PDF-text-layer-grounded pipeline), SFT followed by RLVR with dimension-specific rewards for table reconstruction, chart extraction, and formatting. The report includes 24 pages of data/training pipeline detail for anyone building their own vision models.
Related event: LlamaIndex Releases OCRia-1, Setting New SOTA in Document Parsing(2 posts)→
More from Models
- Blogger predicts Anthropic will ship a new model pre-IPO to maximize its $2T valuation — kimmonismus · 2026-10-26
- MemTensor's MemT-1.0-Nano trends on Hugging Face, a memory-tuned Qwen3.6-35B-A3B finetune — MemTensor · 2026-10-26
- Kimi apologizes for K3.1 Pro quota bugs, gives all subscribers a free 25% quota reset — Kimi_Moonshot · 2026-10-26
- Leak: Fable 5.5 reportedly rolled out to JP Morgan, Meta and other top enterprise clients — ChrisGPT · 2026-10-26
- Subliminal learning transfers hidden traits across models without a shared base model — burny_tech · 2026-10-26
- Jeff Ladish: labs can't train models not to lie, and incentives now push toward subtle deception — JeffLadish · 2026-10-26