What it is
An 8B native unified multimodal model that handles visual understanding, generation, editing, and interleaved creation in one encoder-free, VAE-free system. Instead of separate vision encoders and VAEs, it uses one native pixel-space model with spatially joint patch reconstruction and a specialize-then-unify post-training scheme via RL experts plus on-policy distillation.
Key results
- Supports native generation at up to 4096×4096 resolution
- Compresses each 32×32 image region into a single visual token
- Uses an 8.2B-parameter Mixture-of-Transformers for understanding and generation
A shared native visual representation can transfer multimodal understanding into visual planning, editing, and generation within one end-to-end model.