Update 2026-04-27:

We are glad to share that we have released Tuna-2, a new encoder-free unified multimodal model that does not depend on any pretrained VAE or representation encoder components. Please check out these resources if you are interested!

• Project Website
• Paper link
• Code

We have also released Tuna's inference code in the tuna-2 repo. We are currently working on model checkpoint release for both Tuna and Tuna-2.