27B AI Model Runs on iPhone: PrismML's Bonsai Breakthrough

A 27-billion-parameter AI model is now running directly on an iPhone. PrismML's new Bonsai 27B model operates locally on an iPhone 17 Pro Max, generating about 11 tokens per second. What's interesting is how they achieved this. Bonsai 27B uses an end-to-end low-bit architecture, compressing language weights to roughly 1.125 bits per weight. This reduces the model's size from about 54 gigabytes to just 3.9 gigabytes, while retaining 89.5% of the full-precision model’s average benchmark score. There's also a ternary version, taking up 5.9 gigabytes, which retains 94.6% of the performance. Bonsai 27B is a multimodal model, derived from Qwen3.6-27B, designed for reasoning, coding, and planning. The optional vision component uses a separate 0.63 gigabyte package. The bottom line is that this allows advanced AI to run on a phone, keeping user data on the device rather than sending it to a remote server.