Surya OCR 2.5 ships: 689M-param model tops OmniParseBench at 4.5 pages/s on one H100
pmttyji · reddit · 2026-10-31
Datalab released Surya OCR 2.5, a 689M-param document-focused OCR model:
- Accuracy: 85.9 on OmniParseBench (highest among open models under 1B they've scored) and 83.7% on olmOCR-bench
- Speed: 4.5 pages/s on a single H100 at 192 DPI, 1.7x faster than Surya 2 via MTP
- Multilingual: 74.9% on an internal 91-language benchmark
- Layout analysis with reading order; tables output as HTML with colspan/rowspan merged cells
Code is Apache 2.0; weights use a modified AI Pubs Open Rail-M license—free for research, personal use, and startups under $5M funding/revenue, paid licensing beyond that. Limitations: document OCR only (not natural scenes), table cell geometry requires a Surya 2 checkpoint, and layout/OCR/table recognition need a vllm or llama.cpp backend.
More from Models
- Anthropic kills the 5-hour limit and renames Max 20x to Ultra; users suspect a nerf is coming — doodlestein · 2026-10-31
- Cartesia's cascaded voice agents top the Speech Agent Arena, beating native speech-to-speech — rohanpaul_ai · 2026-10-31
- Correction: newer model gets worse on the extreme-hard tail, pass@5 crossover predicted — teortaxesTex · 2026-10-31
- Hyperflow H3 hits #5 on I2V and #3 on video editing at under half Seedance 2.5's price — ArtificialAnlys · 2026-10-31
- Dev publishes FLARE-based Qwen 3.8 Flash Next images with 1M context for RTX 6000 PRO — johannes_bertens · 2026-10-31
- Rumored GPT-6.5 'Bel' sparks proposal for a no-rules game test over rehearsed launch demos — VraserX · 2026-10-31