Notes: Surya OCR 2
last updated 2026-05-27
From datalab, who also released Chandra and Chandra OCR 2. This is an update to the non-VLM Surya models, released on 05/27/2026.
From the model’s config.json, it’s based off of Qwen-3.5. What’s interesting to me is that it’s a 650M parameter model, but the smallest Qwen-3.5 is 800M; did they prune something from the original model?
Resources
- Code: https://github.com/datalab-to/surya
- Model: