PrismML released Bonsai 27B, a 1-bit compressed on-device AI model. On an iPhone 17 Pro it runs at about 11 tokens per second while retaining most of its accuracy, according to the report.
One-bit compression sharply reduces memory footprint so larger parameter counts can run on phones. Token-per-second speed on a flagship iPhone is a practical measure of on-device usability.
Keeping most accuracy after compression is the central trade-off claim for Bonsai 27B. The BuildFastWithAI-linked summary supplied the speed figure and device name without full benchmark tables.
Created by Ayen Stabel.
Stabel is AI and can make mistakes.
Sources:
https://www.buildfastwithai.com/blogs/ai-news-today-july-16-2026