In a hardware benchmark conducted by Jose Antonio Lanz at Decrypt, the physical limits of mobile computing were just rewritten. PrismML’s new Bonsai 27B model successfully runs full reasoning artificial intelligence locally on an iPhone. By compressing a sprawling 27-billion-parameter architecture down to a 3.9-gigabyte footprint, Bonsai becomes the first model in this capability tier to clear a consumer smartphone’s memory ceiling.
Historically, a 27B-class AI requires approximately 54 gigabytes of memory just to run at half-precision. That physical reality previously restricted these systems strictly to server racks and cloud infrastructure. PrismML shattered this limitation using extreme 1-bit quantization techniques.
How the Compression Works
Bonsai is built on the Qwen3.6 27B base architecture. To shrink it, engineers applied 1-bit and 1.58-bit ternary quantization alongside FP16 group-wise scaling. This highly specific mathematical process prevents the catastrophic accuracy collapse typically seen when conventional models are compressed below four bits.
The results are incredibly fast for a mobile device. The 1-bit binary model runs natively at 11 tokens per second on an iPhone 17 Pro Max. Because iOS restricts individual applications from accessing the device’s total RAM, the 3.9-gigabyte size is critical. It leaves enough functional headroom within the iPhone’s 6-gigabyte app allocation for the system’s KV cache and activations to process data without crashing.
For users with more power, PrismML built a slightly larger 5.9-gigabyte ternary variant. That version achieves 26 tokens per second on an Apple M5 Pro laptop while retaining 94.6% of its full-precision benchmark performance. Both variants are available right now under the Apache 2.0 license.
Hardware Titans Take Notice
The speed and energy efficiency of this offline capability are actively turning heads in Cupertino. Apple is currently evaluating PrismML’s architecture for potential native integration into its own upcoming mobile hardware. PrismML is already targeting a compressed Gemma model as its next project.
Why It Matters
This release fundamentally crosses the threshold separating cloud-based AI from edge computing. Advanced agentic workflows like multi-step planning, coding, and logical debugging no longer require a continuous internet connection or expensive subscription APIs. By processing everything locally on the silicon inside your pocket, Bonsai delivers elite artificial intelligence with absolute data privacy.
The Verdict
PrismML just proved that local, high-parameter AI is entirely viable on consumer smartphones. If you are a developer or an enthusiast looking to build private, offline applications, the Bonsai 27B model is an essential, professional-grade download that pushes Apple silicon to its absolute limit.