SenseTime open-sources SenseNova U1 for Chinese chips, betting on speed

China's SenseTime has released SenseNova U1, an open multimodal model, and is betting on speed rather than maximum size. The main idea behind the release is to teach the model to understand and generate images directly, at lower computational cost and with support for Chinese chips.

Faster without intermediaries

Conventional multimodal systems often work as a pipeline: one block sees the image, another turns it into a text description, a third reasons in words, and then a separate module assembles the visual result again. SenseTime says U1 works differently. In the new NEO-Unify architecture, images and text are processed in a single representation space, without unnecessary intermediate translation steps. As a result, the model responds faster, uses less compute and better preserves meaning and visual detail.

For SenseTime this is not just engineering optimization. The company says outright that U1's key advantage is inference speed. By its own assessment, the model produces results noticeably faster than most open alternatives, and in a number of scenarios approaches the quality of commercial Chinese systems such as Qwen-Image 2.0 Pro and Seedream 4.5. It does not yet match leaders such as GPT-Image-2.0. But its compact size makes U1 potentially suitable not only for data centers but also for PCs or even mobile devices.

• Natively understands images without mandatory translation into text
• Speeds up generation and visual reasoning
• Lowers compute requirements
• Better preserves the structure of complex infographics and text in the frame
• Suited to more compact deployment

A bet on domestic chips

The most politically and commercially important point in the release is compatibility with Chinese hardware. According to SenseTime co-founder and chief scientist Dahua Lin, several Chinese manufacturers have already optimized their accelerators for U1. On launch day, ten local chip designers, including Cambricon and Biren Technology, also announced support for the model.

For the Chinese AI market this is not a minor detail but a question of survival and scale: US export restrictions still complicate access to the most powerful Western chips, above all from Nvidia. SenseTime does not hide that the best foreign accelerators remain useful for the fastest iteration. But the direction is clear: the more models that can be trained and run on local hardware, the lower the dependence on outside suppliers and political risk.

This matters especially for tasks that need fast, real-time visual interpretation of the world. The company ties U1 not only to image generation but also to future robotic systems that need to see a scene, understand spatial relationships and make decisions quickly.

Why open the model

For SenseTime, this launch is also an attempt to regain visibility in the new AI hierarchy. The company grew up on computer vision and facial recognition, but in the era of large language models it ended up in the shadow of younger players such as DeepSeek and MiniMax. Now it is betting on open source: U1 is available for free on GitHub and Hugging Face, and the official release stresses that the U1 Lite series comes in two configurations: a dense 8B-MoT and an A3B-MoT version based on a mixture of experts.

"It is not openness itself that wins, but the speed of iteration," is how

SenseTime explains its new course.

The open release has several goals at once. First, it speeds up feedback from researchers and developers, which helps fix weak points faster and expand use cases. Second, it lets the company maintain international research ties despite sanctions pressure. SenseTime has been under US sanctions for several years over allegations that its technology was used in surveillance systems targeting Uyghurs and other minorities in Xinjiang; the company denies these allegations. Against this backdrop, an open model becomes not only a product but also a tool for a technological and reputational reset.

What this means

The SenseNova U1 release shows how Chinese AI companies are adapting to constraints not only through new models but through a different engineering logic. The focus is shifting from a simple parameter race to efficiency, inference speed, ecosystem openness and compatibility with local hardware.

If this approach works, the winners will not necessarily be the largest models, but those that are deployed faster, cost less to run and fit better into real products, from image generation to robotics.