SenseTime, a Chinese AI institution champion known for its facial designation technology, released a caller unfastened root exemplary connected Tuesday that it claims tin some make and construe images acold faster than apical models developed by US competitors. SenseNova U1 could assistance the institution reclaim mislaid crushed aft it slipped from its spot among the starring players successful China’s AI improvement race.
The model’s concealed condiment is its quality to “read” images without translating them to substance first, speeding up the process and reducing the magnitude of computing powerfulness required. “The model’s full reasoning process is nary longer constricted to text. It tin crushed with images arsenic well,” Dahua Lin, cofounder and main idiosyncratic astatine SenseTime, said successful an interrogation with WIRED.
Lin, who is besides a prof of accusation engineering astatine the Chinese University of Hong Kong, says that models susceptible of processing images straight volition alteration robots to amended recognize the carnal satellite successful the future.
Like DeepSeek's latest flagship model, SenseTime says U1 tin beryllium powered by Chinese-made chips. “Several Chinese home chipmakers person finished optimizing compatibility with our caller model,” Lin says. On merchandise day, 10 Chinese spot designers, including Cambricon and Biren Technology, announced their hardware supports U1.
That flexibility matters due to the fact that US export controls restrict Chinese firms from accessing the world's astir precocious AI chips, peculiarly those utilized for training, which astatine this constituent are chiefly developed by Western companies similar Nvidia. “We volition proceed to propulsion for grooming connected much antithetic chips,” Lin says. But helium besides acknowledges that SenseTime “may inactive request to usage the champion chips to guarantee the velocity of our iteration.”
SenseTime released U1 for escaped connected Hugging Face and GitHub, different motion of however Chinese companies are becoming immoderate of the astir progressive contributors to unfastened root AI.
SenseTime was founded successful 2014 and became a satellite person successful machine vision, which is utilized successful applications similar facial designation and autonomous driving. But erstwhile ChatGPT and different AI systems powered by earthy connection processing became the hottest happening successful the tech industry, SenseTime began struggling to crook a nett and fell down newer Chinese startups similar DeepSeek and MiniMax.
SenseTime says it hopes that releasing SenseNova-U1 publically for anyone to usage volition assistance it drawback up with some home and Western AI players. Lin says the institution yet made the determination past twelvemonth to absorption connected unfastened root due to the fact that of the adjuvant feedback it gets from researchers, which enables the institution to iterate faster. “In this time and age, being unfastened root oregon closed root is not the winning factor; the velocity of iteration is,” Lin explains.
Going unfastened root besides helps SenseTime proceed collaborating with planetary researchers without the interference of geopolitics. The institution has been sanctioned repeatedly by the US authorities successful caller years implicit allegations that its facial designation exertion helped powerfulness surveillance systems utilized to show and detain Uyghurs and different number groups successful China’s Xinjiang region. As a result, US firms are restricted from investing successful SenseTime and selling definite technologies to it without a license. (SenseTime has denied the allegations.)

A illustration representation created utilizing SenseNova U1. Generated utilizing AI
Seeing Clearly
In an accompanying method report, SenseTime claims that SenseNova-U1 generates higher-quality images than each different unfastened root models presently connected the market. Its show is comparable to starring Chinese closed root models similar Alibaba’s Qwen and ByteDance’s Seedream, but it inactive lags down manufacture leaders similar GPT-Image-2.0, which came retired conscionable a week ago.
But the model’s main selling constituent is its quality to make images overmuch faster than each of those models. It relies connected an innovative method operation called NEO-Unify that SenseTime previewed earlier this year.











English (CA) ·
English (US) ·
Spanish (MX) ·