SenseTime SenseNova U1.5 Brings 8B-MoT Native Unified Vision With Open Training Code

SenseTime has announced the launch of SenseNova U1.5, an 8-billion-parameter model that integrates understanding, reasoning, and image generation into a unified multimodal system. This model enhances spatial reconstruction accuracy by processing text and images within a shared architecture. The design improvements include replacing independent patch MLP decoding with a lightweight spatial decoder, allowing visual tokens to be reshaped into a two-dimensional feature field. This enables neighboring regions to exchange information before finalizing pixel processing, supporting native generation up to 4K while reducing seams and texture discontinuities.