Apple is in early negotiations with PrismML, a small AI startup backed by Khosla Ventures. The talks aim to explore technology that would allow significantly larger AI models to run entirely on an iPhone without cloud servers.
PrismML specializes in compressing massive AI models. It has compressed Alibaba's 27-billion-parameter Qwen 3.6 model from roughly 54GB of memory to under 4GB. This makes it small enough to run directly on an iPhone 17 Pro.
The company uses a technique called ultra-low-bit weight compression. This method reduces the precision of a model's parameters while preserving performance. Venture capitalist Vinod Khosla, whose firm led PrismML's $16.25 million seed round, has called the approach a "mathematical breakthrough."
PrismML plans to open-source its compressed model on July 14, 2026. The startup's ambitions extend to eventually compressing trillion-parameter models for edge deployment.
For Apple, this fits a broader strategy. The tech giant's current on-device model, AFM 3 Core Advanced, is sparsely activated. The company has previously acquired AI firm DarwinAI and held talks with Google and OpenAI to bolster its capabilities. These latest discussions coincide with iOS 27 updates focused on enhancing Siri and Apple Intelligence.
The technology's potential to reshape where AI runs makes it significant. Once the compression technology becomes open-source, the race to deploy it will expand beyond Apple.