Datalab benchmarks 17 open OCR models on OmniParseBench; Chandra OCR 2 tops at 90.1
VikParuchuri · x · 2026-10-25
Datalab (Vik Paruchuri) ran 17 open-weight OCR models on OmniParseBench — 16,288 pass/fail tests over 2,937 pages, building on AllenAI's olmOCR-bench — each with the vendor's documented serving setup:
- Overall: Chandra OCR 2 (4.5B, Datalab's own) leads at 90.1, ahead of dots.mocr (3.0B, 87.9), MinerU2.5-Pro (1.2B, 87.1) and Infinity-Parser2-Pro (35.1B MoE, 87.0).
- Category winners: MinerU2.5-Pro is best at tables (92.1); Infinity-Parser2-Flash best at math (96.9); Chandra OCR 2 sweeps text, layout, handwriting, tiny text, multi-column and non-English. Models without layout output get no headline score.
- Method notes: parameter counts reflect inference weights (multi-token-prediction heads excluded); all scores are published — including Datalab's own models losing in places — and every adapter plus the benchmark code (Apache-2.0) is on GitHub for auditing.
Related event: Datalab Launches OmniParseBench, Chandra OCR 2 Tops Open-Source Ranking(2 posts)→
More from Models
- Claude Code sends 16,581 tokens per step yet scores equal a 100-line mini-harness — bendee983 · 2026-10-25
- Developers vent frustration at Claude's 'deluded sense of moral superiority' — arthurcolle · 2026-10-25
- vLLM matches DeepInfer on DeepSeek-V4.1-Flash across batch sizes 8–256 — vllm_project · 2026-10-25
- I measured $20 coding plans: ~33 tasks per 5h on Sonnet, only ~6 on Opus — Sufficient-Storage87 · 2026-10-25
- Yacine: after DeepSeek, open the weights and 'trust in humanity' — IgorCarron · 2026-10-25
- Tech veteran with 30 years in the field asks: when did knowledge become intelligence? — SuzannahB1001 · 2026-10-25