Startup PrismML has compressed the 54GB Qwen 3.6 model — a 27 billion parameter system — down to just 4GB without degrading performance, according to a report from The Information. The breakthrough targets a core constraint in mobile AI: server-scale models typically exceed the memory and compute budget of smartphones, forcing Apple to rely on smaller, less capable on-device models.

Apple has reportedly held meetings with PrismML to explore integrating its compression technology into future iPhone hardware. The company's preference for on-device inference stems from privacy, latency, and offline functionality goals, but current iPhone models top out at a few billion parameters. PrismML's approach, described as mathematical optimization rather than quantization alone, preserves the complex reasoning and autonomous agent capabilities that larger parameter counts enable.

No partnership is confirmed, and no timeline exists for consumer availability. If adopted, the technology could let Siri and third-party apps run sophisticated chat, reasoning, and agent workflows locally — a shift that would reduce cloud dependency and differentiate Apple's AI stack from rivals still leaning on server-side processing.