Deepseek-VL2

DeepSeek made waves with their R1 language model, but their multimodal capabilities (especially image understanding) are not good enough: But, they are rapidly evolving. DeepSeek-VL2, their new open-source family of Mixture-of-Experts (MoE) vision-language models, is a big step forward, achieving strong performance with a much smaller activated parameter count, thanks to its MoE design. And the exciting news: there is a new Hugging Face Spaces demo – you can now try these models without needing to deploy heavily (normally you would need more than 80GB of GPU resources, which is almost impossible for most of us) So check it out and see what DeepSeek brings next to suprise everyone :)

Awesome, that's good news, I did not like the "No text extracted" error when i tried to upload images of any kind...

WoW! The speed at which DeepSeek is evolving is really mindblowing. Congrats on the launch and sending wins to the team :)

Great models from DeepSeek! Evolving fast, great value for us!

For transparency in performance you should add Qwen 2.5 VL 7B as well

Deepseek's greatest value is that R1 cost nothing compared to Chatgpt o1; There are still no good use cases for "middle detail" vision models yet (Good enough to identify things in pictures but not good enough to drive a car etc.) Awesome release though.

All the best for the launch @zac_zuo!