PrismML has unveiled Bonsai 27B, a groundbreaking model based on the open-source Qwen3.6 27B, marking the first instance of such a large AI model functioning on a smartphone. The original Qwen3.6 27B model typically occupies around 54 GB of storage, making it unfeasible for mobile devices. However, the development team successfully compressed the model down to 3.9 GB while maintaining a minimal loss in quality. This was achieved by training the model with a strictly confined set of values from the beginning, which allows it to retain about 90-95% of the performance found in the full-sized version. In practical applications, this AI can engage in intricate conversations, write code, analyze documents and images, and perform complex tasks autonomously. The startup was established by a group from Caltech and has received backing from prominent investors including Khosla Ventures, Google, and Samsung, raising an impressive $16.25 million in its initial funding round. Reports indicate that Apple has initiated discussions with PrismML regarding the potential integration of this technology. Apple champions the concept of on-device AI to ensure user data remains secure and off the cloud, although some of the most advanced functionalities of the updated Siri still rely on remote servers. Bonsai 27B is now available for download on HuggingFace. Are you interested in having powerful AI capabilities directly on your phone?
Informational material. 18+.
