Bonsai 27B Runs Locally on iPhone - A 27B Model in 3.9GB
Bonsai 27B model is developed with 1-bit quantization for local iPhone use.
PrismML has developed Bonsai on top of Qwen3.6-27B by quantizing the weights to 1-bit, reducing the size from 54GB to 3.9GB. This allows it to run on a phone while retaining about 90% of benchmark scores, averaging 76.1 across 15 benchmarks.