PrismML's 1-bit Bonsai 27B, a 27B-class model that runs on an iPhone, draws developer attention

PrismML’s Bonsai 27B, a 27.3-billion-parameter multimodal model that runs on an iPhone, drew a wave of developer attention following its July 14 release.

Built on Alibaba’s Qwen3.6-27B with a 262,000-token context window, Bonsai 27B ships in a 1-bit binary variant of 3.9GB, roughly 14 times smaller than the 54GB FP16 version, and a 1.58-bit ternary variant of 5.9GB. That compression is the point: it shrinks a model class that normally needs a server down to something that fits on a phone.

PrismML says the 1-bit variant scores 76.11 on average across 15 benchmarks in thinking mode, about 89.5% of the full-precision baseline, with 91.66 on math and 81.88 on coding tasks. Those numbers come from PrismML and have not been independently verified.

On speed, the company says the model runs at 11 tokens per second on an iPhone 17 Pro, 44.2 on an Apple M5 Pro and 66.4 on an Apple M5 Max. The weights are free to download under an Apache 2.0 license on Hugging Face, where the model card lists the variants.

Aggressive quantization to 1-bit weights typically trades some accuracy for size, and the vendor’s own benchmark scores measure that trade only on PrismML’s chosen tests. The interest suggests developers are testing whether the claims hold on real hardware rather than benchmark suites.