PrismML has launched the Bonsai 27B, the first AI model with 27 billion parameters designed to operate on consumer smartphones. The model, compressed to 3.9 GB, runs at 11 tokens per second on an iPhone 17 Pro Max. A ternary version, compressed to 5.9 GB, achieves 26 tokens per second on an M5 Pro laptop. This compression, based on Caltech technology, reduces model weights to binary values, making the model 14 times smaller than its full-precision counterpart. The ternary Bonsai 27B scores 80.49 on average across 15 benchmarks, maintaining 94.6% of the full model's performance. Released under the Apache 2.0 license, PrismML is in discussions with Apple to further develop this technology.