Indra OCR Fact-Check: Accuracy Claims Hold, But the 17x Speed Claim Is Really ~3x
TheOyinbooke · x · 2026-10-15
A developer ran Datalab's own 2,937-page benchmark on a Mac Studio to test the 1M-parameter Indra OCR model. Accuracy claims hold: 28.5 pages/second on 18 CPU cores, within 0.2 points of official numbers, with full layout and line bounding boxes. But the speed claim doesn't: Indra was 3x faster than Tesseract on his Mac, not the 17x advertised. Full comparison against Tesseract and PaddleOCR included.
More from Models
- First Proof batch 3: mathematicians grade GPT-6, Opus 5.5 and Kimi K3 on research-level proofs — littmath · 2026-10-15
- Google tops Signal65 PINNACLE agentic leaderboard: Gemini 4 Argon at $0.87 per correct task beats Sonnet and GPT — ryanshrout · 2026-10-15
- DeepSeek V4.1 Flash tops cost-efficiency at $0.134 per task, leaderboard shows — bindureddy · 2026-10-15
- Whatever Happened to Test-Time Scaling? Problems Needing pass@8192 Now Solve One-Shot — marc_lelarge · 2026-10-15
- cantrell on Gemini 4 Argon: heavy thinker, building long decode to push past 64K token limit — cantrell · 2026-10-15
- Opus 5.5 diagnoses fake OneDrive 'desktop takeover' in one prompt, maps 3 autostart paths — rehtorical · 2026-10-15