Meet Bonsai: The First 27B AI Model That Fits on Your Phone
PrismML released Bonsai 27B, a 27-billion-parameter model compressed enough to run on consumer hardware, including an iPhone 17 Pro Max. It uses an end-to-end ternary/binary scheme that reduces weights to -1, 0, or +1, with shared scaling factors, making the binary version about 14× smaller than full precision. Unlike typical quantized models, it keeps embeddings, attention, and the output head compressed too. Despite the extreme compression, quality remains close to the original: the ternary model averages 80.49 on 15 benchmarks, about 94.6% of full precision, while the 1-bit version scores 76.11. It performs strongly on math, coding, knowledge, and tool use, and its hybrid attention design enables a practical 262K-token context window on-device. PrismML also provides speculative decoding for faster GPU inference, and the models are free under Apache 2.0. Apple is reportedly evaluating the compression technology for possible on-device use.
