Namo-V2 release

[lucasjinreal]

Hello! We just finished training of Namo-500M-V2!

The v2 version boosted a lot in OCR ability, and it adopted SiglipV2 as vision encoder. Moreover, it can handle native resolution input as usual in the Namo series.

We are still conducting mDPO on the trained V2 model. We are exploring if mDPO could further enhance the tiny model's ability or not. Stay tuned for our final release.

Also, the new V2 version will have two versions, one with 578 tokens per / img, and one with 256 tokens per / img.

[forensicmike]

Cool! Just sharing this as well, some research that shows how GRPO can be used to enhance things like bounding box IoU detection (and includes some new datasets):
https://github.com/Liuziyu77/Visual-RFT
Research paper:
https://arxiv.org/abs/2503.01785

[lucasjinreal]

Oh, good to see it, DoD (detection on description) and Odd on the roadmap of our ultimate goal. Will adopt these wonderful work in the near future.