Apple talks about AI start-up to let big models run directly on iPhone

On July 15, according to CNBC, Apple is in preliminary contact with AI startup PrismML and evaluating its AI model compression technology, hoping to allow large AI models to run directly on the iPhone without relying on cloud computing. PrismML claims that it can reduce model memory usage by 10 to 15 times, allowing models with 27 billion parameters to run on iPhone15 and above devices, while improving response speed by 6 to 8 times and reducing energy consumption by 3 to 6 times. If the technology matures, Apple is expected to transfer more Siri and generative AI functions to local operation, further improving response speed, privacy protection capabilities and offline experience, and reducing cloud computing costs. However, industry insiders pointed out that this technology still needs to be verified by large-scale practical applications, and its impact on battery life, stability and long-term performance remains to be seen.