SenseTime Unveils Open-Source Image Model Optimised for Domestic Hardware Amid Export Restrictions

Sanctioned Chinese artificial intelligence firm SenseTime has released SenseNova U1, an open-source image generation and interpretation model specifically engineered to operate on domestic hardware. The launch represents a strategic response to United States export controls that restrict access to advanced Western chips, particularly those used for training developed by companies like Nvidia. By partnering with local manufacturers, SenseTime aims to circumvent these geopolitical constraints while maintaining competitive performance in the rapidly evolving AI sector.

The model distinguishes itself through a native image processing architecture that bypasses the traditional step of translating images into text. This technical innovation allows the system to reason directly with visual data, resulting in significantly faster generation speeds and reduced computing power requirements compared to top Western competitor models. Dahua Lin, cofounder and chief scientist at SenseTime, noted that this approach enables robots to better understand the physical world without the latency associated with converting visual inputs into textual descriptions first.

On the day of the release, ten Chinese domestic chip designers confirmed hardware compatibility with the new model. Among the confirmed partners are Cambricon and Biren Technology, both key players in the local semiconductor industry. This flexibility is critical as it allows the firm to continue iterating on its technology without relying on restricted foreign hardware, although Lin acknowledged that the company may still require the best available chips to ensure optimal iteration speeds in the short term.

SenseTime made the model freely available on Hugging Face and GitHub, marking a deliberate pivot toward open-source development to accelerate feedback loops and global collaboration. The firm claims that SenseNova U1 generates higher-quality images than any other currently available open-source model. However, the company admits that its performance still lags behind recent industry leaders such as GPT-Image-2.0, which was released just a week prior to this announcement.

Founded in 2014, SenseTime was previously a world leader in computer vision and facial recognition but has struggled to maintain profitability and market share against newer startups following the rise of natural language processing models. The firm has faced repeated US sanctions over allegations that its facial recognition technology supported surveillance systems in China's Xinjiang region, allegations the company has consistently denied.

Looking ahead, SenseTime is actively collaborating with ACE Robotics, a startup led by another SenseTime co-founder, to apply this technology to humanoid robots and geospatial simulations. The ability to process visual information natively is expected to allow robots to act faster and make fewer mistakes in complex environments, potentially driving the current boom in humanoid robotics within the region.