SenseTime · 2026-09-10 · notable
SenseNova-U1.5 report — the recipe behind SenseTime's 8B unified model
SenseNova-U1.5 is SenseTime's 8B-MoT unified multimodal model, and its technical report is now public. The paper covers an encoder-free, VAE-free design that reads and generates images at native resolutions up to 4K.
SenseTime publishes how it trained an 8B model that reads and draws images in one stack.
What is it?
The technical report for SenseNova-U1.5 is now public, covering the architecture, training recipe and evaluation results behind SenseTime's 8B-MoT unified multimodal model. The weights have been on Hugging Face since 20 August 2026 under Apache 2.0, but this paper is the first full write-up of how they were made. It reached second place on Hugging Face's trending papers the day after it posted.
How does it work?
No separate vision encoder and no VAE sit in the SenseNova-U1.5 pipeline — a single stack handles both understanding and generation. The visual interface is trained with spatially coherent patch reconstruction, and generation runs at native resolutions up to 4K. After pretraining, the team tunes separate experts for visual aesthetics, bilingual text rendering, infographic generation and image editing, then folds them back together with multi-expert on-policy distillation.
Why does it matter?
Most models that both read and create images bolt a generator onto a language model. The SenseNova-U1.5 authors argue the end-to-end route works, reporting gains in image fidelity, text rendering, complex composition and multi-reference editing while keeping subject identity, geometry and unmodified regions intact. The team says it will open-source the training code, covering supervised fine-tuning, reinforcement learning and on-policy distillation.
Who is it for?
multimodal researchers
Try it
https://huggingface.co/sensenova/SenseNova-U1.5-8B-MoT