SenseTime released SenseNova U1.5, an 8-billion-parameter Mixture-of-Transformers native unified vision model that treats understanding, reasoning, and pixel-space generation as one native multimodal system without external encoder or VAE modules, mapping pixels and text into a shared backbone. It replaces independent patch MLP decoding with a lightweight spatial decoder that reshapes visual tokens into a two-dimensional feature field restored through Pixel Shuffle stages and local convolutions, while still mapping each 32-by-32 region to one visual token and supporting native generation up to 4K with fewer seam and texture discontinuities than the prior U1 design, aided by resolution-aware noise conditioning across wider aspect ratios. Post-training follows a specialize-then-unify recipe: separate reinforcement-learning experts target visual aesthetics, bilingual text rendering, infographic layout, and image editing, then multi-expert on-policy distillation folds those skills into a single student. SenseTime reports stronger results on open image-generation and editing suites versus U1, including competitive bilingual text rendering and multi-reference editing, while keeping multimodal understanding scores broadly intact on standard STEM, VQA, and OCR benchmarks. Alongside the technical report on arXiv and Hugging Face Papers, SenseTime is open-sourcing training code covering supervised fine-tuning, reinforcement learning, and on-policy distillation, with model assets pointed to SenseNova and Hugging Face collections under the OpenSenseNova organization. The release is the full U1.5 product stack