SenseNova U1 Open-Sourced

SenseTime unveiled SenseNova U1, combining understanding, reasoning, and generation into a single unified multimodal model.

• The core architecture NEO-Unify removes the visual encoder(VE) and VAE, processing images and text in the same representation space.
• The open-source release consists of two versions in the U1 Lite lineup: 8B-MoT(dense) and A3B-MoT(MoE).
• Official materials emphasize strengths in image understanding, image generation/editing, complex infographic generation, and image-text continuous generation, along with top-tier performance and faster inference compared to open-source peers.
• Weights and inference code have been released on GitHub and Hugging Face, with a larger-parameter version also announced.