Bonsai 27B (1-bit)

1-bit/ternary quantized rebuild of Qwen3.6-27B by PrismML — compresses the 27B model to ~3.9GB while retaining ~90% of original quality; first 27B-class model to run interactively on a phone (iPhone 17 Pro Max) and at ~44 tok/s on a consumer laptop. Apache 2.0, open weights + code + technical report.

CONTRIBUTED BY