Meet Bonsai: The First 27B AI Model That Fits on Your Phone

PrismML released Bonsai 27B, a 27-billion-parameter AI model compressed to 3.9 gigabytes that runs on an iPhone 17 Pro Max at 11 tokens per second, the company announced earlier this week. The binary variant fits within a smartphone memory budget while a ternary version at 5.9 gigabytes reaches about 26 tokens per second on an M5 Pro laptop. Both are available free under Apache 2.0. The compression method, based on Caltech intellectual property, reduces each weight to a single sign with a shared 16-bit scaling factor per 128 weights, achieving 1.125 bits per weight for the binary model and 1.71 bits for the ternary model without higher-precision escape hatches for embeddings or attention layers. Across 15 benchmarks in thinking mode, ternary Bonsai 27B averages 94.6% of full-precision performance. The model uses a hybrid attention backbone with roughly 75% linear layers, enabling a 262,000-token context window on device. PrismML CEO Babak Hassibi confirmed to CNBC that Apple is in early talks evaluating the compression technology for potential on-device use. Hassibi said a compressed Gemma model is next in the pipeline followed by larger frontier models.