Apple explores PrismML tech to run 27‑billion‑parameter AI model on iPhone

Apple is in early discussions with the AI startup PrismML about using its model‑compression technology for future iPhones. PrismML claims to have reduced Alibaba’s open‑source Qwen 3.6 language model—from 54 GB and 27 billion parameters to about 4 GB—without a noticeable loss in performance. If the claim holds, the compressed model could run fully on the iPhone 17 Pro, allowing all 27 billion parameters to be active at once, unlike Apple’s current on‑device model (AFM 3 Core Advanced) that contains 20 billion parameters but only activates 1–4 billion at a time. Running such a large model on‑device would cut reliance on Apple’s Private Cloud Compute, lower operating costs and improve user privacy by keeping data on the handset. Apple has not confirmed any deal and the talks remain at a preliminary stage, but the prospect highlights the company’s continued focus on expanding on‑device AI capabilities.

The reported negotiations follow Apple’s broader strategy of compressing AI models to fit within the memory and processing limits of its mobile hardware while delivering richer Siri and Apple Intelligence features.