OpenBMB open-sources MiniCPM OCR family: 1.3B model tops open-source OCR benchmarks, 2B evaluator hits 0.936 AUROC
AdinaYakup · x · 2026-10-28
OpenBMB released its first open-source OCR models under Apache 2.0. MiniCPM OCR 1.0 (1.3B) delivers 55% higher processing efficiency with DSpark and handles complex layouts, tables, and formulas; its chart shows 91.3 average across four benchmarks, beating Xiaomi-OCR-0, PaddleOCR-VL-1.6, and GLM-OCR at 1.1 pages/s. MiniCPM OCR Jev (2B) evaluates OCR accuracy including other models' outputs, scoring 0.936 AUROC on JevOCRBench — its 93.6 overall beats Qwen3.5-35B and Gemma 4-31B despite the tiny size.
Related event: OpenBMB Open-Sources MiniCPM-OCR Family with 1.3B SOTA Model(6 posts)→
More from Models
- Open-weight models now hold 72% of OpenRouter tokens as Tim Shi bets enterprises will own their intelligence — thursdai_pod · 2026-10-28
- Google launches Genie 4: its most realistic interactive world model with native audio and real-time steering — BorisMPower · 2026-10-28
- EmbeddingGemma 2 runs real-time embeddings on a MacBook CPU alone: 20ms text, 100ms images — generativist · 2026-10-28
- Google Gemma rounds up EmbeddingGemma 2 community projects, from WebGPU photo search to Apple Silicon — GlennCameronjr · 2026-10-28
- DeepSeek says "I'm ChatGPT" in identity test, resurfacing distillation allegations — ResonantFork · 2026-10-28
- Developer chats with a lamp fully locally using the new LFM3-Audio model — ngxson · 2026-10-28