Paper Notes: OvisOCR2
last updated 2026-08-25
OvisOCR2 is a Qwen-3.5 0.8B finetune that currently (as of 08/25/2026) holds the number one position on the OmniDocBench leaderboard. As noted in the paper, it’s the first single-shot VLM to hold this position. Typically, the best performing models have been pipelined VLMs like MinerU2.5 and PaddleOCR-VL (check my series on OCR’s Cambrian Explosion to learn more about the difference between pipelined and single-shot models). As a bonus, the model weights are Apache 2.0 licensed!
Perhaps more importantly, I’ve seen real people use it and get great results!
So the question is: how is a 0.8B model so strong?
As with most things in machine learning, it’s not really the model. Alibaba built a robust data engine, that relies on a lot of manual filtering.
A Robust Data Engine
Distillation Pipeline
The second training trick of OvisOCR2 is the distillation pipeline. They:
- Supervised finetune (SFT) both Qwen-3.5 0.8B and Qwen3.5 4B on the same dataset
- RL tune Qwen3.5 4B, because they found that RL’ing the 0.8B was too high variance and did not converge
- perform On-Policy Distillation (OPD) from the 4B to the 0.8B
By doing this,