Apple Eyes AI Compression Startup That Ran a 27B-Parameter Model on an iPhone 17 Pro

The dream of running sophisticated AI models entirely on-device — no cloud round trips, no latency, no data leaving your phone — has been just out of reach. The models that do the most interesting work are simply too large to fit in a phone’s memory budget. That may be changing.

Apple is in talks with PrismML, a Caltech spin-off whose core technology compresses large language models to roughly 1/14 of their original size — a reduction that could let the iPhone run models that currently require data center GPUs. The Information reports that Apple is evaluating PrismML’s approach for future iPhones.

PrismML’s method is not traditional quantization, where weights are squeezed into fewer bits at the expense of precision. Instead, it uses a native 1-bit representation: weights are stored as either -1 or +1, with grouped scaling factors that preserve dynamic range. The company claims the technique maintains accuracy close to the original FP16 model while delivering up to 8x faster inference and cutting energy consumption by 75-80%. Memory use drops by more than 90%.

The startup demonstrated the technology by taking Alibaba’s open-source Qwen 3.6 — a 27-billion-parameter model — and running it entirely on an iPhone 17 Pro. The model operated without any cloud fallback, a proof point Apple’s hardware and AI teams would find hard to ignore.

For Apple, the appeal is straightforward. On-device AI means faster response times, better privacy, and lower operational costs compared with cloud-reliant architectures. The company has been investing heavily in its own silicon and neural engines, but the frontier of what’s possible on-device has been limited by model size. A compression technique that cuts a 27B-parameter model down to something that fits comfortably inside an iPhone’s memory changes that calculus.