According to the tech media The Information, Apple is in talks with AI startup PrismML to evaluate the feasibility of running larger-scale AI models directly on iPhones. This move indicates that Apple may achieve a major breakthrough in on-device AI capabilities, allowing smartphones to run large language models that previously required cloud computing power.

PrismML is an AI startup spun off from Caltech, with its core breakthrough being native 1-bit model compression technology. This technology can compress the model size to about one-fourteenth of the full-precision version, reducing memory usage by over 90%. Traditional quantization schemes usually only reduce precision but still retain multi-bit weights, while PrismML's weights are represented by only -1 and +1, combined with group scaling factors for computation, completely redefining the model's storage and inference methods at the architecture level.

Apple, iPhone14, Launch Event

No "high-precision escape route," yet close to original precision

PrismML claims its technology does not include the "high-precision escape route" commonly found in traditional quantization schemes — a compromise where some key layers retain high precision to compensate for performance loss. Under the complete use of 1-bit weights, this technology can still maintain precision levels close to FP16 models. At the same time, the inference speed can be increased up to eight times, and energy consumption can be reduced by 75% to 80%, meaning that the computing power and power requirements for running large AI models on mobile devices have been significantly lowered without sacrificing model quality.

27B parameter Qwen 3.6 has been fully operational on the iPhone 17 Pro

The most critical practical evidence has now emerged. PrismML compressed the large language model Qwen 3.6, which is open-sourced by Alibaba, and successfully ran it fully on the iPhone 17 Pro. A 27B parameter model would almost certainly be impossible to run smoothly on a smartphone under traditional schemes, but after being compressed by PrismML's 1-bit method, it not only ran successfully but also maintained near-original inference quality. Apple was drawn to this quantization capability and hopes to further enhance the reasoning performance of local AI models with it.

For Apple, the strength of on-device AI capabilities directly affects the competitiveness of the Apple Intelligence ecosystem. Currently, the scale and functionality of AI models on iPhones are still limited by memory and power consumption. If PrismML's compression technology can be implemented, the iPhone could run larger-scale models without increasing hardware costs, achieving more complex multi-turn dialogues, image understanding, and intelligent agent task orchestration. When flagship-level large models can also run on smartphones, the balance between on-device AI and cloud AI might be reset.