PrismML has announced Bonsai 27B, the first 27B-class model capable of running on a smartphone. The model achieves this by using 1-bit and ternary weights to reduce its size to 3.9 GB and 5.9 GB respectively.
Kalshi is launching prediction markets for AI computing power, allowing users to track future GPU rental prices through forward curves extending up to a year. Meanwhile, Google announced Gemma 4 optimized for the Pixel 10's TPU, supporting offline tasks like audio transcription and device control.
The article also highlights the WANDR benchmark for evaluating research agents and notes that open weights now handle the majority of production tokens.