Apple is reportedly in early negotiations with PrismML, a small AI startup backed by Khosla Ventures, to explore technology that could allow significantly larger AI models to run entirely on an iPhone. No cloud servers, no data leaving your pocket.
What PrismML actually does
PrismML’s headline trick is compressing Alibaba’s Qwen 3.6 model, a 27-billion-parameter beast that normally requires roughly 54GB of memory, down to under 4GB. That makes it small enough to run actively on an iPhone 17 Pro without phoning home to external servers.
The company uses what it describes as ultra-low-bit weight compression, a technique that reduces the precision of each parameter’s numerical representation while preserving the model’s overall performance. Vinod Khosla, the venture capital legend whose firm led PrismML’s $16.25 million seed round in early 2026, called the approach a “mathematical breakthrough.”
PrismML has announced plans to open-source its compressed model on July 14, 2026. The startup is already looking beyond current capabilities, with ambitions to eventually compress trillion-parameter models for edge deployment.
