Datalab OCR beats Azure and AWS Textract by wide margins on non-Latin scripts
VikParuchuri · x · 2026-10-28
New OmniParseBench numbers shared by VikParuchuri: across 897 non-Latin text tests on 297 pages, Datalab lite scores 76.3 vs Google Document AI 71.9, Azure layout 33.9, and AWS Textract just 2.5 — sometimes beating other providers by a wide margin on multilingual documents. Hard pages can use balanced or accurate mode; a playground is available to try it directly.
Related event: Datalab launches lite OCR at a fraction of cloud rivals' cost(4 posts)→
More from Models
- New LALA benchmark: 236 humanities questions suggest more post-training can make LLMs dumber — JasonBotterill · 2026-10-28
- Liquid AI's LFM3 puts browser agents, voice, and vision on enterprise-controlled devices — JosephJacks_ · 2026-10-28
- Liquid AI ships LFM3-1.7B, a 1.7B conversational model built for edge deployment — JosephJacks_ · 2026-10-28
- Zero-training PCA head lifts 577M EmbeddingGemma-2 to #1 on MTEB audio, beating 9B models — JeanMichelRanu · 2026-10-28
- Arabic course creator weighs Opus 5.5 against discounted Gemini 4 Argon at $6 — Various_Specific_623 · 2026-10-28
- Google bans user from Gemini, AI Studio and Search after repeated safety-filter flags — immellocker · 2026-10-28