Surya OCR 2.5 cuts repetitive-loop failures 38x vs Surya 2, hitting token cap on just 0.1% of requests
VikParuchuri · x · 2026-10-30
Repetitive output loops are a common failure in small OCR models. Datalab reports that on a hard set, Surya OCR 2.5 hits the token cap (looping) on only 0.1% of requests (2 of 1,390), versus 3.8% for Surya 2 + MTP (38 of 988) — a 38x reduction. Tested on one H100, vLLM 0.25.1, 192 DPI, pooled over 1/8/32/128 concurrent requests.
Related event: Surya OCR 2.5 released as top sub-1B open-source OCR(6 posts)→
More from Models
- Reddit post accuses Cohere of fear-mongering about Qwen security to time its multibillion-dollar raise — yuicebox · 2026-10-30
- Heretic v2: uncensored Qwen 3.8 125B quant tuned for reverse engineering with activation-energy bit allocation — jjusko20 · 2026-10-30
- Critic Ed Zitron admits he hasn't used this year's models, sparking AGI debate — iruletheworldmo · 2026-10-30
- Heretic-v2 quant of Qwen3 Swift 1.5 Flash ships with reverse-engineering-tuned imatrix for 96GB VRAM — jjusko20 · 2026-10-30
- Altman: a new OpenAI model 'knocked down' research problems, Navier-Stokes proof cost ~$10M — chrismattmann · 2026-10-30
- xAI has everything it needs, so why is it falling behind at the frontier? — Snoo26837 · 2026-10-30