SenseTime, a Chinese language AI firm greatest recognized for its facial recognition know-how, launched a brand new open supply mannequin on Tuesday that it claims can each generate and interpret pictures far quicker than high fashions developed by US opponents. SenseNova U1 might assist the corporate reclaim misplaced floor after it slipped from its place amongst the leading players in China’s AI growth race.
The mannequin’s secret sauce is its potential to “learn” pictures with out translating them to textual content first, dashing up the method and lowering the quantity of computing energy required. “The mannequin’s total reasoning course of is not restricted to textual content. It may possibly cause with pictures as effectively,” Dahua Lin, cofounder and chief scientist at SenseTime, mentioned in an interview with WIRED.
Lin, who can be a professor of data engineering on the Chinese language College of Hong Kong, says that fashions able to processing pictures immediately will allow robots to higher perceive the bodily world sooner or later.
Like DeepSeek’s newest flagship mannequin, SenseTime says U1 could be powered by Chinese language-made chips. “A number of Chinese language home chipmakers have completed optimizing compatibility with our new mannequin,” Lin says. On launch day, 10 Chinese language chip designers, together with Cambricon and Biren Know-how, introduced their {hardware} helps U1.
That flexibility issues as a result of US export controls prohibit Chinese language corporations from accessing the world’s most superior AI chips, notably these used for coaching, which at this level are primarily developed by Western firms like Nvidia. “We are going to proceed to push for coaching on extra completely different chips,” Lin says. However he additionally acknowledges that SenseTime “should want to make use of the very best chips to make sure the velocity of our iteration.”
SenseTime launched U1 without spending a dime on Hugging Face and GitHub, one other signal of how Chinese language firms have gotten a few of the most lively contributors to open supply AI.
SenseTime was based in 2014 and have become a world chief in pc imaginative and prescient, which is utilized in functions like facial recognition and autonomous driving. However when ChatGPT and different AI techniques powered by pure language processing grew to become the most popular factor within the tech business, SenseTime started struggling to show a revenue and fell behind newer Chinese language startups like DeepSeek and MiniMax.
SenseTime says it hopes that releasing SenseNova-U1 publicly for anybody to make use of will assist it meet up with each home and Western AI gamers. Lin says the corporate lastly made the choice final yr to concentrate on open supply due to the useful suggestions it will get from researchers, which allows the corporate to iterate quicker. “Nowadays, being open supply or closed supply just isn’t the profitable issue; the velocity of iteration is,” Lin explains.
Going open supply additionally helps SenseTime proceed collaborating with worldwide researchers with out the interference of geopolitics. The corporate has been sanctioned repeatedly by the US authorities lately over allegations that its facial recognition know-how helped energy surveillance techniques used to watch and detain Uyghurs and different minority teams in China’s Xinjiang area. Because of this, US corporations are restricted from investing in SenseTime and promoting sure applied sciences to it with out a license. (SenseTime has denied the allegations.)
Seeing Clearly
In an accompanying technical report, SenseTime claims that SenseNova-U1 generates higher-quality pictures than all different open supply fashions presently available on the market. Its efficiency is akin to main Chinese language closed supply fashions like Alibaba’s Qwen and ByteDance’s Seedream, but it surely nonetheless lags behind business leaders like GPT-Picture-2.0, which got here out only a week in the past.
However the mannequin’s important promoting level is its potential to generate pictures a lot quicker than all of these fashions. It depends on an modern technical construction known as NEO-Unify that SenseTime previewed earlier this yr.
