Would We Adopt the Newly Released Lightweight Multimodal Model Nanonets-OCR-s?

[xiyuan27]

This model（Nanonets-OCR-s） has just been released and has already reached 201,530 downloads in just a few days. It requires only 8 GB of VRAM to run and is reported to perform very well.

Nanonets-OCR-s: https://huggingface.co/nanonets/Nanonets-OCR-s/tree/main

In addition, there are other impressive models, such as Ovis2-4B and InternVL3-8B. Rankings and scores can be found here:
https://rank.opencompass.org.cn/leaderboard-multimodal/?m=REALTIME

[myhloli]

We utilize the VLM model from opendatalab/MinerU2.0-2505-0.9B to achieve end-to-end output. Thanks to its optimized model architecture and reduced parameter count, it offers advantages in resource utilization and parsing speed.

[GeorgeDeac]

I have also thought of this, if it would be possible to make it work with multiple VLM backbones.