SenseTime released SenseNova U1 on Tuesday, an open-source model that generates and interprets images and is engineered to run on Chinese-made silicon. On launch day, 10 Chinese chip designers, including Cambricon and Biren Technology, said their hardware supports the model. The release is SenseTime's clearest attempt yet to climb back into China's top tier of AI labs after slipping behind DeepSeek and MiniMax.
The pitch is speed. SenseTime claims U1 generates images faster than every leading rival, including Alibaba's Qwen and ByteDance's Seedream, though it still trails OpenAI's GPT-Image-2.0, which shipped a week earlier, on raw quality. The model is small enough to run on PCs and phones, and SenseTime has posted weights to both Hugging Face and GitHub.
The architectural bet is called NEO-Unify, a structure SenseTime previewed earlier in 2026. It lets the model reason over images natively rather than transcribing them to text first, which the company says cuts both latency and compute draw. "The model's entire reasoning process is no longer limited to text. It can reason with images as well," Dahua Lin, SenseTime's cofounder and chief scientist, told Wired.
Key facts
- 01SenseTime released SenseNova U1 on Tuesday as an open-source image generation and reasoning model on Hugging Face and GitHub.
- 0210 Chinese chip designers, including Cambricon and Biren Technology, confirmed hardware compatibility on release day.
- 03SenseTime, founded in 2014, says U1 trails GPT-Image-2.0 on quality but generates images faster than Alibaba's Qwen and ByteDance's Seedream.
- 04U1 uses a NEO-Unify architecture, previewed earlier in 2026, that lets the model reason over images directly without converting them to text first.
- 05The company decided last year to pivot to open source after falling behind newer Chinese rivals DeepSeek and MiniMax.
Chip independence is the other half of the story. US export controls block Chinese firms from buying Nvidia's most advanced training accelerators, and SenseTime itself has been sanctioned repeatedly over allegations that its facial recognition tools fed surveillance of Uyghurs in Xinjiang. SenseTime denies the allegations, but the practical result is the same: US firms cannot invest in it or sell it certain technologies without a license.
“Ten Chinese chip designers, including Cambricon and Biren Technology, announced day-one support for SenseNova U1, a workaround to US controls on Nvidia's top training silicon.”— Jaeden Schafer
Like DeepSeek's flagship release, U1 is positioned as a model that runs well on domestic chips. "Several Chinese domestic chipmakers have finished optimizing compatibility with our new model," Lin said. He added that SenseTime will keep pushing training across more chip families, while conceding the company "may still need to use the best chips to ensure the speed of our iteration."
SenseTime, founded in 2014, built its name on computer vision for facial recognition and autonomous driving. The pivot to large language and multimodal models has been bumpy. Once a leader in Chinese AI, it has struggled to turn a profit since ChatGPT redefined what investors and customers expect from the category, and newer rivals have raced past it.
Lin said the company decided last year to commit to open source, reasoning that external researchers would surface bugs and improvements faster than an internal team could. "In this day and age, being open source or closed source is not the winning factor; the speed of iteration is," he said. Open releases also let SenseTime keep collaborating with researchers abroad despite the sanctions perimeter around it.
Adina Yakefu, an AI researcher at Hugging Face, said the architecture is what sets U1 apart from the field. "This is a more ambitious approach, as it still faces significant practical challenges," she said. "It's good that they decided to open source it, so the community can explore and test it more widely."
Lin sees robotics as the natural target. A robot operating in a cluttered environment has to parse an enormous amount of visual information in real time, and converting every frame into a text description is expensive and lossy. "It has to think, 'how should I deal with all the clutter in this room? If there is a complicated machine in front of me, which button should I press?'" he said. SenseTime does not build robots itself but is working with ACE Robotics, a startup run by another SenseTime cofounder.
The caveats are real. U1 still lags GPT-Image-2.0 on output quality, NEO-Unify is unproven at scale outside SenseTime's own benchmarks, and "compatible with Chinese chips" is not the same as "trained end-to-end without Nvidia" — Lin himself signaled the company will keep using top-tier accelerators where it can. Yakefu's framing — ambitious but with significant practical challenges — is the honest read.
Still, the release fits a pattern that should worry anyone counting on export controls to keep a lid on Chinese model quality. DeepSeek, Qwen, Seedream and now SenseNova U1 are all shipping as open weights, all claiming workable performance on domestic silicon, and all iterating in public faster than most US labs are willing to. The competitive question for OpenAI, Google and Anthropic is no longer whether Chinese models are good enough — it is whether the open-source distribution model coming out of Beijing, Hangzhou and Shenzhen ends up setting the default for how multimodal AI gets deployed everywhere outside the US.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
