Apple's on-device AI path may have advanced after startup PrismML reportedly compressed Alibaba's Qwen 3.6 model to run on the iPhone 17 Pro, according to The Information. PrismML reduced the 27 billion-parameter model from about 54GB to less than 4GB, claiming the reduction does not hurt performance, with all 27 billion parameters active at once to handle complex chat, reasoning, autonomous agents and software coding. Apple has held meetings with PrismML about possible uses of its technology, as the startup's approach could appeal to Apple's emphasis on running more AI features on devices rather than relying fully on cloud data centers. The report highlights Apple's push to close the AI gap without matching the data-center spending arms race led by Microsoft, Amazon, Meta and Google. The next test is whether PrismML's compression method can scale beyond a technical milestone into real Apple product features.