SenseNova U1

What it is

SenseNova U1 is an open-weight multimodal model from SenseTime: a single model that both reads images and produces them. Its initial weights and inference code were released on 27 April 2026 under the Apache License 2.0.^[1] "Open-weight" means the trained parameters themselves are published, so anyone can download, run and modify the model instead of reaching it only through a company's interface; Apache 2.0 is a permissive licence that allows commercial reuse.

An architecture without an encoder or a VAE

Most multimodal systems are assembled from separate parts. A vision encoder turns an image into a sequence of feature vectors that a language model can read; when the system draws an image, a decoder such as a VAE (variational autoencoder) squeezes pixels into a compact latent space and expands them back into pixels. Each component is trained on its own objective and only then attached to the language model.

U1 is built on SenseTime's NEO-unify architecture, which removes both the visual encoder and the VAE and models language and vision end to end.^[1] Images enter and leave the model in much the same way as text tokens, and the whole stack is trained as one system rather than a chain of separately trained modules. The practical point is that understanding and generation share one backbone, so what the model learns while reading images is available to it when it draws.

Underneath, the model uses what SenseTime calls mixture-of-transformers backbones.^[1]

Two backbones

SenseNova-U1-8B-MoT is a dense model, with about 8 billion parameters for understanding and about 8 billion for generation. SenseNova-U1-A3B-MoT instead uses a mixture of experts.^[1] A mixture of experts keeps many separate parameter blocks and routes each piece of input through only a few of them, so the model can hold more parameters in total than it activates for any given token.

What it can do

On the understanding side the model takes up to 32K tokens of context, enough for a long stretch of interleaved text and images. On the generation side the default output resolution is 2048x2048.^[1] Both capabilities come from the same set of weights rather than from two models talking to each other.

Release history and ecosystem

The April release was the 8B dense model. The technical report and the A3B weights followed on 10 May 2026, and more specialised versions arrived over the following weeks: infographic-variant weights on 15 May 2026 and an Interleaved variant on 11 June 2026.^[1]

Support was added to stable-diffusion.cpp and llama.cpp on 26 October 2026.^[1] Both are open-source projects that run generative models in C and C++ on ordinary hardware, so support there means the model can be executed outside SenseTime's own inference stack.

SenseTime said the U1 series passed 11,000 combined GitHub stars less than four months after being open-sourced, and that it was adapted to more than 10 domestic chip platforms on day zero.^[2]

Where it sits in the series

SenseTime describes the U1 line as comprising SenseNova U1, U1.5 and U1.5-Lite.^[2] The line began with the first open weights in April 2026 and widened from there: a technical report, additional and specialised variants, and third-party inference support.^[1]