Surya OCR 2.5 released as top sub-1B open-source OCR
Datalab (author VikParuchuri) has released the open-source OCR model Surya OCR 2.5. At just 689M parameters, it scores 85.9 on OmniParseBench, making it the most accurate open-source OCR model under 1B parameters; it runs 1.7x faster than Surya 2, reaching 4.5 pages per second on an H100. The model is hosted on Hugging Face, the code is on GitHub, with an accompanying technical blog—ready to try or deploy.
Confirmed
- 689M parameters, 85.9 on OmniParseBench, positioned as the strongest open-source OCR under 1B
- Inference 1.7x faster than the previous generation, 4.5 pages per second on an H100
- Over 80% accuracy across 52 languages on internal benchmarks, retaining strong multilingual capability
- Dedicated training cleanup for the repetition-loop problem common in small OCR models: on a hard test set, only 0.1% of requests hit the token limit (just 2 of 1390 requests), with repetition loops reduced 38x versus the previous generation
- Uses a character-level tokenizer augmented with about 20,000 common multi-character units, for a vocabulary of 88,526 tokens covering emoji, fancy math letters, and common words, yielding fewer tokens on the same text
- While congratulating @hheiden12 on a new release, VikParuchuri teased that more updates are coming for both Chandra and Surya
Why it matters
- Achieving balance across accuracy, speed, and stability with under 1B parameters lowers the barrier to local deployment and large-scale document processing—a clear net gain for the open-source document parsing ecosystem.
2026-10-30 ~ 2026-10-30 · 6 related posts
Primary sources
- Surya OCR 2.5 launches: most accurate open OCR under 1B params, 4.5 pages/s on H100 — VikParuchuri ·
- Surya OCR 2.5 cuts repetitive-loop failures 38x vs Surya 2, hitting token cap on just 0.1% of requests — VikParuchuri ·
- Surya OCR 2.5: model weights, code, and blog post all publicly available — VikParuchuri ·
- [source] Surya OCR 2.5 launches: most accurate open OCR under 1B params, 4.5 pages/s on H100 — VikParuchuri · 2026-10-30
- [source] Surya OCR 2.5: model weights, code, and blog post all publicly available — VikParuchuri · 2026-10-30
- Character-level tokens plus common words: inside Surya OCR 2.5's 88,526-token vocabulary — VikParuchuri · 2026-10-30
- [source] Surya OCR 2.5 cuts repetitive-loop failures 38x vs Surya 2, hitting token cap on just 0.1% of requests — VikParuchuri · 2026-10-30
- Surya OCR 2.5 hits 80%+ accuracy in 52 languages, loops 38x less than v2 — VikParuchuri · 2026-10-30
- Surya OCR 2.5 scores over 80% on 52 languages in internal benchmark — VikParuchuri · 2026-10-30