Apple has long championed the idea that the future of artificial intelligence (AI) lies not in the cloud, but on local devices. This approach predates the current boom in generative AI; the company has consistently opted to run its machine learning (ML) algorithms on its own hardware for privacy reasons. While Siri AI has incorporated some cloud-based functionalities, local models like AFM Core and AFM Core Advanced remain central. Recently, Apple customized an ML model featuring 20 billion parameters for their iPhone 17 Pro, 17 Pro Max, and iPhone Air, which was widely considered a significant achievement. However, a startup named PrismML has raised the bar even higher. Reports indicate that this emerging firm has successfully adapted Qwen 3.6, an open-source model from Alibaba, to run with 27 billion parameters on an iPhone 17 Pro.
The Future of AI: Local Processing
According to The Information, this breakthrough enables complex chat response tasks, reasoning challenges, and even fully autonomous coding agents. The public will gain access to this model shortly, as PrismML plans to make it available this upcoming Tuesday. The company anticipates that within the next three years, up to 95% of the required AI capabilities will be processed locally on smartphones, laptops, and other devices. Baba Hassibi, the CEO of PrismML, predicts that only 5% of high-end tasks will require cloud resources in the near future, asserting that this is the direction the industry must embrace.
The Shift from Cloud to Local Devices
This shift could cause concern among AI investors, who have invested hundreds of billions into data centers equipped with powerful GPUs. If PrismML’s vision becomes a reality, the inference – or output generated based on prompts – would transition to local devices, marking a substantial shift in AI deployment. However, it’s crucial to note that training models will likely still require data centers for the foreseeable future.
Initial Talks with Apple
Founded as a spinoff of the California Institute of Technology (Caltech), PrismML employs “mathematical tricks” to compress Qwen 3.6 significantly, reducing its size from 54 GB to just 4 GB without quality loss. Caltech holds the necessary patents for these techniques. Apple has already introduced advanced technology for AFM Core that utilizes flash storage to manage the RAM limitations of large models, allowing only essential components to be loaded as needed.
Currently, Apple is on the lookout for additional AI firms to acquire, although it’s unclear if PrismML will be among them. However, reports indicate initial conversations have taken place regarding the potential use of PrismML’s technology within Apple’s ecosystem. Despite this, cloud models still hold a distinct advantage: they can be updated rapidly and automatically, removing the need for users to take action to access the latest versions.
Conclusion
The developments from PrismML signify a pivotal moment in AI technology, particularly regarding local processing capabilities. As the lines between local and cloud computing continue to blur, users may allow for an era of more robust, privacy-centric AI applications right in their pockets. With local devices taking on more AI responsibilities, the landscape of machine learning is set to undergo transformative changes that will redefine user experiences.
