SenseNova-U1.5 unifies 8B multimodal model with native 4K output

SenseNova has released SenseNova-U1.5, an 8.2B-parameter model its authors describe as an "8B-MoT native unified multimodal model that understands, reasons about, and generates visual content within an encoder-free and VAE-free architecture." The paper reports GenEval 0.92, CVTG-2K text-rendering 0.948 and ImgEdit 4.59, and native generation up to 4096×4096.

The architectural bet is Mixture-of-Transformers: understanding and generation share attention as a cross-stream interface, but keep separate attention projections, normalization layers and feedforward modules. Training runs across five stages ending in what the authors call "multi-expert on-policy distillation" — four specialized experts for aesthetics, OCR text rendering, editing and infographics, then consolidation into one model over 800 optimizer steps.

On the interleaved benchmark VBVR-Pro-Bench, the paper claims 67.6% in-domain and 48.7% out-of-domain, and says the model "Outperforms leading proprietary systems (Nano-Banana-Pro, GPT-Image-2)." The abstract argues the broader point plainly: "these findings position native unified modelling as a promising path towards systems that perceive, reason and create within a fully end-to-end framework."

The release promise is unusually complete. "We will open-source training code, including supervised fine-tuning, reinforcement learning, and on-policy distillation," the authors write, alongside checkpoints on a Hugging Face collection. It arrives inside a run of Chinese open-weight releases we've been tracking closely on the China AI page, and lands the same week two small-model papers from other Chinese labs crossed our open-source tracker. One was LUMIA Lab's NCP-ArchPreview, an 8.9B latent-space LM matching OLMo-3-7B on 51% of tokens.

The paper does not publish inference cost or latency at 4K, and the VBVR-Pro-Bench comparison against Nano-Banana-Pro and GPT-Image-2 is self-reported.