Apple is intensifying its efforts to streamline large AI models for local operation on iPhones, a strategy aimed at reducing cloud computing costs and enhancing user privacy protection. Meanwhile, a small startup that recently emerged from stealth mode has announced a significant achievement: successfully deploying a model on an iPhone with a parameter count far exceeding any previous mobile AI model.
The startup, named PrismML, claims to have compressed the open-source large language model Qwen 3.6 from Chinese internet giant Alibaba to run locally on an iPhone 17 Pro. This model boasts 27 billion parameters — analogous to synapses in the human brain, which determine a model's ability to handle complex data. In contrast, the vast majority of local models on mobile devices can typically only activate a few billion parameters at a time.
Trillion-parameter super-large models remain incapable of running on mobile devices. However, PrismML states that this 27-billion-parameter dense model, now operational on an iPhone, can perform complex tasks such as sophisticated dialogue, deep logical reasoning, fully automated intelligent agents, and code development. The open-source version of this model will be available for download next Tuesday.
This previously undisclosed technical milestone reflects a broader industry trend: shifting AI computation from expensive data center servers to local execution on end-user devices. Tech giants like Microsoft, Amazon, and Meta have invested hundreds of billions of dollars in building computing power centers to meet anticipated surges in AI demand.
Apple, however, has not followed suit by investing heavily in massive computing clusters. Instead, it has publicly advocated for running as many AI functions as possible locally on the iPhone, rather than in the cloud. Apple believes that on-device AI better fulfills its commitments to user privacy and security.
In an interview, PrismML's CEO, Babak Hassibi, predicted that the vast majority of AI computation will eventually be performed locally on end-user devices.
"Imagine the scenario three years from now: 95% of intelligent computing will be done locally on your phone, laptop, or smart home devices, with only 5% of tasks requiring extremely high computing power being offloaded to the cloud. The industry widely favors this development path," Hassibi stated.
Hassibi added that lightweight, local deployment of models "fundamentally reshapes the cost structure of the AI industry."
Several leading AI investment firms are optimistic about this startup's technical approach. Khosla Ventures, an early investor in OpenAI, participated in PrismML's $16.25 million seed funding round earlier this year. The firm's founder, Vinod Khosla, stated in an interview that PrismML's technology represents a "fundamental breakthrough," which led to the decision to invest.
"In 2018, we invested in OpenAI, betting heavily on the Transformer architecture. Now, the industry needs entirely new approaches to building AI, and our team has been actively seeking out such innovative solutions," Khosla said.
PrismML relies on a proprietary mathematical algorithm to compress Qwen 3.6 to a fraction of its original size. While conventional model compression techniques often lead to performance degradation, the company claims its in-house lightweight solution does not compromise model effectiveness. They have reduced the original ~54GB Qwen 3.6 model to under 4GB.
PrismML originated from the California Institute of Technology. CEO Hassibi is a professor of electrical engineering at the university, where he and his co-founders developed the core mathematical algorithm. Caltech holds all patents for the technology and has granted PrismML an exclusive license for commercial use.
Hassibi revealed that PrismML plans to continue compressing even larger models, with the goal of reaching trillion-parameter levels, aiming to compete with top-tier flagship models like OpenAI's GPT and Anthropic's Claude.
PrismML's technological solution is highly attractive to Apple. At its Worldwide Developers Conference in June, Apple announced that the long-awaited major overhaul of Siri would rely on Google's Gemini model. Due to the large size of the model, Siri's advanced features will still require computing power from Google Cloud's Nvidia chips.
Simultaneously, Apple announced that several new AI features for the iPhone would shift to local operation. Apple's own in-house on-device model, which uses a sparse architecture with 20 billion parameters, only activates between 1 and 4 billion parameters at a time. In contrast, PrismML's local model can fully and simultaneously activate all 27 billion parameters.
Previous reports from The Information indicated that Apple encountered bottlenecks when trying to compress its own large models for the iPhone, resulting in significant performance degradation.
There are also reports that Apple is seeking to acquire companies that can strengthen its on-device AI capabilities and has engaged in technical cooperation discussions with PrismML.
Other technical approaches for on-device AI exist within the industry. For example, startup Argmax focuses on performing local preprocessing of voice and image data before uploading the processed data to the cloud for complex reasoning.
A major reason for the popularity of such hybrid edge-cloud solutions is the extremely rapid iteration speed of cloud-based large models, which are updated weekly. Proponents of hybrid architectures argue that models running entirely locally cannot benefit from the latest cutting-edge iterations available in the cloud.
Comments