PrismML在高通骁龙峰会展示智能眼镜专用小型语言模型

AI实验室PrismML在高通骁龙峰会上推介了为智能眼镜设计的1-bit Bonsai语言模型1。该模型拥有20亿参数，针对视觉和语言任务进行了优化，能够在高通Snapdragon AR1 Gen 1平台上本地运行1。

PrismML声称其压缩技术可将大型模型压缩4倍，同时保留原有性能1。这一成果致力于推进可在边缘设备上运行的开放权重人工智能1。PrismML由加州理工研究人员创立，由伯克利的Ion Stoica担任顾问1。不过目前尚无搭载该技术的商用智能眼镜产品发布1。

AI laboratory PrismML showcased its 1-bit Bonsai LLM at Qualcomm's Snapdragon Summit, bringing a specialized language model to Qualcomm Snapdragon AR1 Gen 1-powered smart glasses.1 The model, featuring 2 billion parameters, has been optimized to run visual and language tasks directly on device rather than relying on cloud processing.1

PrismML claims its compression technology can reduce large models to one-quarter their original size while maintaining performance capabilities.1 The company, founded by researchers from the California Institute of Technology and advised by Ion Stoica of Berkeley, is focused on advancing open-weight AI systems that operate locally on hardware.1 While the technology was demonstrated at Wednesday's Snapdragon Summit, no smart glasses products incorporating PrismML's models have been released to market as of now.1