China's BenZhi Activation Leads Charge in Full-Stack On-Device AI Innovation

In response, BenZhi Activation introduces a disruptive, native on-device approach that rebuilds the AI software and hardware stack from the ground up. This strategy bypasses traditional model compression methods, enabling powerful AI models to run efficiently and securely on end-user hardware without sacrificing performance.

TMTPOST -- BenZhi Activation, an AI startup incubated by the Institute of Parallel and Distributed Systems (IPADS) at Shanghai Jiao Tong University, is pioneering a full-stack native on-device AI solution designed to transform personal computing.

Founded and led by Associate Professor Ze-Yu Mi, the team draws on IPADS’s global leadership in operating and distributed systems research, backed by world-class expertise in large-scale on-device AI models and infrastructure.

Their portfolio includes internationally recognized open-source projects like PowerInfer and SmallThinker, demonstrating strength in both cutting-edge AI development and engineering implementation.

As AI technology reshapes industries, BenZhi Activation aims to revolutionize how billions of devices—including PCs, smartphones, and smart terminals—process data by embedding advanced AI capabilities locally, rather than relying on cloud-based models. Cloud AI currently faces significant challenges around privacy, latency, and personalization.

Uploading sensitive personal data to cloud servers raises security concerns and discomfort among users reluctant to entrust their “digital lives” to third parties. Additionally, cloud interactions incur high costs and latency that limit frequent, in-depth AI use. Finally, generalized cloud models struggle to deeply personalize experiences, lacking the ability to learn continuously from individual user data while maintaining privacy.

In response, BenZhi Activation introduces a disruptive, native on-device approach that rebuilds the AI software and hardware stack from the ground up. This strategy bypasses traditional model compression methods, enabling powerful AI models to run efficiently and securely on end-user hardware without sacrificing performance.

The company has achieved collaborative breakthroughs across on-device large model algorithms, infrastructure systems, and hardware optimization, delivering AI that is both highly capable and fully private.

BenZhi Activation has already delivered world-leading innovations. In December 2023, they released PowerInfer, an on-device AI infrastructure capable of running tens-of-billions-parameter models on consumer NVIDIA GTX 4090 GPUs with performance near that of data-center A100 GPUs, boasting inference speeds over 11 times faster than prior methods. This open-source project quickly rose to the top of GitHub’s global trending list. By June 2024, their PowerInfer-2 system, incorporating a proprietary TurboSparse sparsification method, enabled a 4.7-billion-parameter model to run smoothly on smartphones, outpacing the international benchmark llama.cpp by 29 times and marking a leap from desktop to mobile deployment at scale.

Looking ahead, BenZhi Activation will collaborate with Shanghai Jiao Tong University to release the world’s first batch of native large AI models pre-trained and architected explicitly for edge deployment. These models address the tight computational, memory, and storage constraints of low-cost hardware, allowing seamless operation of tens-of-billions-parameter models on devices costing only a few hundred yuan. This milestone exemplifies the team’s full-stack capabilities, from foundational algorithm design to on-device infrastructure. Their earlier release of SmallThinker, a 3-billion-parameter reasoning model optimized for on-device use, achieved over 100,000 downloads on HuggingFace within a week and ranked second globally among trending AI models.

Leading venture capitalists praise BenZhi Activation’s innovative approach. Chen Yu, Partner at Yunqi Partners, noted that the company effectively bridges the gap between powerful AI models and mainstream device computing capabilities, accelerating low-cost and efficient AI deployment. Liu Shui, Managing Director at Baidu Ventures, emphasized the significance of coordinated optimization across models, systems, and hardware to reduce AI costs and deliver real-time, private AI experiences on smart devices. Huang Xinxin from Lighthouse Capital highlighted BenZhi Activation’s rare combination of top-tier R&D and production expertise, positioning the startup as a global leader in edge AI innovation.

BenZhi Activation’s pioneering native on-device AI technology signals a fundamental shift in the industry, empowering billions of users with secure, responsive, and deeply personalized AI experiences directly on their devices. This approach is set to redefine the future of personal intelligence, putting advanced AI capabilities squarely in users’ hands.

本文系作者 zhangxinyue 授权钛媒体发表,并经钛媒体编辑,转载请注明出处、作者和本文链接
本内容来源于钛媒体钛度号,文章内容仅供参考、交流、学习,不构成投资建议。
想和千万钛媒体用户分享你的新奇观点和发现,点击这里投稿 。创业或融资寻求报道,点击这里

敬原创,有钛度,得赞赏

赞赏支持
发表评论
0 / 300

根据《网络安全法》实名制要求,请绑定手机号后发表评论

登录后输入评论内容

快报

更多

20:08

波罗的海干散货指数涨1.74%

20:07

台积电、中芯国际等多家半导体厂商回应海外氦气供应短缺影响

20:05

2025年95种网售产品国家监督抽查不合格率为19.1%,同比下降4.4个百分点

19:51

高盛称近期美股涨势若要延续,需要美联储重新转向降息

19:50

教育部发布《中国青少年阅读素养框架》教育行业标准

19:47

佩斯科夫:俄仍愿接收伊朗浓缩铀,但美方未予采纳

19:46

乌外长:乌方愿在土耳其与俄方举行领导人层级会晤

19:46

北证50指数基金规模上限将提升

19:42

瑞达期货:2026年一季度净利同比预增135.64%-165.25%

19:40

北方稀土:2026年目标营收440亿元以上,利润总额35亿元以上

19:39

长盈精密:2025年度净利润5.98亿元,同比下降22.53%

19:38

涉及环境监测、食品检验、机动车检验等,28家机构涉嫌违法违规

19:37

广州海珠“脑机10条”正式发布,最高支持1000万元

19:35

伊朗副外长:伊朗要求彻底结束整个地区的冲突

19:33

美银:2030年全球服务器市场将达1.5万亿美元,AI贡献超八成份额

19:32

广交会爆款订单起飞,中国无人机订单飞涨

19:31

山东印发鲁西崛起五年实施方案,聚焦五市跨越发展

19:26

证监会就《期货公司监督管理办法(征求意见稿)》及配套实施规定公开征求意见

19:23

阿里云:受短信服务综合成本显著上涨影响,5月20日起调整国内短信服务产品价格

19:21

3连板均瑶健康:AKK菌粉销售目前对公司利润影响极小,相关产品未来能否实现规模化市场推广存在重大不确定性

扫描下载App

Baidu
map