OpenBMB ships 1.3B OCR model scoring 93.68 plus a 2B AI judge for OCR errors
Equivalent-Grass-527 · reddit · 2026-10-28
OpenBMB released two open-source models: MiniCPM-OCR 1.0 (1.3B) scores 93.68 on Real5-OmniDocBench, targeting real-world documents like curved pages, phone photos, tilted scans, and complex layouts, with DSpark decoding boosting end-to-end efficiency by 55%.
The more interesting piece is MiniCPM-OCR Jev (2B), an OCR quality judge that compares the original image with candidate text to catch errors in formulas, tables, and text — 0.936 AUROC, close to a 397B model on the same eval. Jev can also grade outputs from other OCR systems, so you don't need to replace your existing pipeline.
Both models are open source, supporting vLLM, SGLang, llama.cpp, and Ollama.
Related event: OpenBMB Open-Sources MiniCPM-OCR: 1.3B Parser Plus 2B Quality Judge(7 posts)→
More from Models
- Letta's FrontierMemory leaderboard: Letta Dream tops at 73.3%, beating OpenAI and Claude agents — afurgs · 2026-10-28
- xAI Rebrands as SpaceXSI, Launches Grok 4.7 and Unified Multimodal API — RachelVT42 · 2026-10-28
- 'Benchmaxxed' Models Sink on Release as Community Waits for Next Attempts — rickasaurus · 2026-10-28
- Artificial Analysis launches Progress Index v1 to track capability since Nov 2022 — ArtificialAnlys · 2026-10-28
- AA v5 tasks consume 59k median output tokens, up two-thirds from v4.3 — ArtificialAnlys · 2026-10-28
- Effort settings now swing scores 20+ points at 10x cost, AA v5 finds — ArtificialAnlys · 2026-10-28