FULL STORY
PixelUMM: NVIDIA's Encoder-Free Unified Multimodal Model Goes Open Source
NVIDIA and the University of Waterloo released PixelUMM, an encoder-free unified multimodal model operating directly in pixel space. Days later, the 15B-parameter model was open-sourced, completing the line from announcement to release.
2026-10-02 ~ 2026-10-06 · 2 episodes · 5 posts
Episode 1 · NVIDIA Releases PixelUMM, Encoder-Free Unified Multimodal Model (2026-10-02, 3 posts)
NVIDIA and the University of Waterloo released PixelUMM, an open-source encoder-free multimodal model built on Qwen3-8B that handles image and video understanding and generation directly in pixel space.
- NVIDIA releases PixelUMM on Hugging Face, a Qwen3-8B-based image-to-text model — nvidia · 2026-10-02
- NVIDIA's PixelUMM: encoder-free unified image and video understanding and generation — nvidia · 2026-10-02
- NVIDIA and Waterloo release PixelUMM: an encoder-free multimodal model reading and writing raw pixels — CSProfKGD · 2026-10-02
Episode 2 · NVIDIA Open-Sources PixelUMM, a 15B Encoder-Free Unified Multimodal Model (2026-10-05, 2 posts)
NVIDIA researchers have open-sourced PixelUMM, a 15B-parameter encoder-free unified multimodal model that handles both image and video understanding and generation directly in raw pixel space, eliminating the need for a VAE or pretrained vision encoder.