PrismML compresses Qwen 3.6-27B to <4GB for iPhone

PrismML has developed a compressed version of Alibaba's open-source Qwen 3.6 model, enabling the 27-billion-parameter large language model to run on an iPhone 17 Pro. The startup reduced the model's size from approximately 54 gigabytes to less than 4 gigabytes using a mathematical technique derived from Caltech research.

• All 27 billion parameters remain active during inference, unlike Apple's new Siri model which activates only 1 to 4 billion of its 20 billion parameters.
• The compressed model supports complex chat, reasoning, autonomous agents, and software coding tasks.
• PrismML claims the compression technique does not hinder performance despite the significant size reduction.
• The open-source model is scheduled for release next Tuesday.

PrismML CEO Babak Hassibi predicts that this approach will fundamentally change the economics of AI by shifting the vast majority of intelligence processing to local devices rather than the cloud.