最新模型
SenseNova-U1.5 Publishes Report on Its Native Unified Vision Architecture, but the Full Training Pipeline Has Yet to Be Released
SenseNova-U1.5 uses a Mixture-of-Transformers (MoT) architecture that does not rely on a separate vision encoder or VAE, enabling a single model to handle understanding, reasoning, generation, and editing. The technical report added on September 11 details its training methods and evaluations. Existing weights and inference code are available, but the SFT, RL, and multi-expert distillation code has not yet been fully released.
On September 11, the SenseNova team published the U1.5 technical report, providing a more complete explanation of the architecture and training behind the weights previously released in August. The model follows the NEO-unify approach: instead of attaching a separate vision encoder or VAE, it uses spatially continuous patch reconstruction to establish a shared interface for pixels and text. The same autoregressive system can perform visual question answering, text-to-image generation, image editing, and interleaved image-text generation. Its Mixture-of-Transformers (MoT) architecture uses different experts to address conflicts between capabilities and supports native output at resolutions up to 4K.
During post-training, separate experts are optimized for aesthetics, Chinese and English text rendering, infographics, and editing, after which their capabilities are merged through multi-expert on-policy distillation. The paper claims improvements in evaluations of image quality, complex composition, multi-reference image editing, and text rendering. The model card provides Apache 2.0-licensed weights, Transformers inference examples, and code for text-to-image generation and editing. The reference environment uses Python 3.11, PyTorch 2.8, and CUDA 12.8. Loading the model requires trust_remote_code=True, and it is not currently hosted directly by a Hugging Face Inference Provider.
The more consequential question is whether eliminating the need to align separate representations for understanding and generation can reduce cross-modal information loss—not how impressive the showcase images look. For now, the reported figures mainly come from the team’s own evaluations. The total parameter information shown on the model card and the 8B-MoT designation also need clearer explanation. Dense small text, mixed Chinese-English typography, exact object counts, hands, and drift across multi-turn edits remain known weaknesses. The team has pledged to release its SFT, reinforcement learning, and distillation pipelines. Until then, third parties can reproduce inference but cannot fully verify the data recipe or training results.