{"id":70170,"date":"2026-07-10T17:47:52","date_gmt":"2026-07-10T15:47:52","guid":{"rendered":"https:\/\/www.apfelpatient.de\/?p=70170"},"modified":"2026-07-10T17:50:03","modified_gmt":"2026-07-10T15:50:03","slug":"apples-larger-ai-models-on-the-iphone","status":"publish","type":"post","link":"https:\/\/www.apfelpatient.de\/en\/rumors\/apple-groessere-ki-modelle-auf-dem-iphone","title":{"rendered":"Apple is testing significantly larger AI models directly on the iPhone"},"content":{"rendered":"

A young startup has reportedly managed to run a language model with 27 billion parameters fully on an iPhone 17 Pro \u2013 larger than Apple's own most powerful on-device model. According to a recent report, Apple has already met with those responsible to explore what the technology could mean for the iPhone and iPad.<\/strong><\/p>\n\n\n\n

Apple's AI strategy has been under pressure for months: Publicly, the company emphasizes processing directly on the device as a promise of data privacy, but in reality, the new Siri generation runs on Google's Gemini technology \u2013 and the most demanding requests are routed to the cloud. Against this backdrop, The Information<\/a> reports that Apple has been in talks with the startup PrismML, whose compression technology is said to run significantly larger models locally on the iPhone. The fact that Apple's own private cloud compute servers were recently described as too weak for modern Frontier models<\/a> makes the appeal of a more powerful on-device solution understandable.<\/p>\n\n\n\n

What the startup aims to achieve<\/h2>\n\n\n\n

PrismML claims to have compressed Qwen 3.6 \u2013 an open-source language model from the Chinese company Alibaba with 27 billion parameters \u2013 so drastically that it runs entirely on an iPhone 17 Pro. According to the company, the model shrank from around 54 gigabytes to under 4 gigabytes without any performance loss. This is made possible by a mathematical method based on extremely efficient 1-bit and ternary weights, which drastically reduces storage requirements.<\/p>\n\n\n\n

The startup, a spin-off from the California Institute of Technology (Caltech), has reportedly raised $16.25 million in seed funding, which also included participation from Khosla Ventures. PrismML plans to release its model as open source on July 14th; it is intended to be able to handle tasks in software development, among other things.<\/p>\n\n\n\n

The crucial difference: all parameters are active simultaneously<\/h2>\n\n\n\n

To understand the appeal for Apple, it helps to look at how it works. The number of parameters roughly describes how complex the relationships are that a model can process \u2013 more parameters generally mean greater performance. The problem is that large models usually exceed the storage and processing power of a smartphone.<\/p>\n\n\n\n

Apple's own most powerful on-device model, AFM 3 Core Advanced, circumvents this limitation through a so-called sparse architecture. While it possesses 20 billion parameters, it only activates 1 to 4 billion of them at any one time \u2013 depending on the task, only a fraction of the model is "awakened" to avoid overloading the iPhone. In iOS 27, this model powers, among other things, the more expressive voices of Siri AI and the improved system-wide dictation function on the iPhone 17 Pro and iPhone Air.<\/p>\n\n\n\n

PrismML's approach reverses this principle: all 27 billion parameters are intended to be permanently active simultaneously. The startup sees this as its most important competitive advantage, because it would give the model its full capacity for complex tasks \u2013 from sophisticated dialogues and logical reasoning to autonomously acting agents.<\/p>\n\n\n\n

Why Apple is interested in this topic<\/h2>\n\n\n\n

Larger processors integrated directly into the device would allow more Apple intelligence functions to be handled locally instead of via Apple's private cloud computing. This would offer Apple two advantages: lower server operating costs and a stronger data privacy argument, since sensitive requests wouldn't even leave the device. This two-tiered approach of local processing and private cloud computing<\/a> has been the foundation of Apple's AI communication for years.<\/p>\n\n\n\n

The need is real: Siri's most demanding functions now run on Gemini models<\/a> in the cloud, while Apple has to expand its own server infrastructure in parallel. A technology that brings significantly more intelligence back to the iPhone would make the company somewhat less dependent on this setup.<\/p>\n\n\n\n

Why skepticism remains appropriate<\/h2>\n\n\n\n

As impressive as the numbers sound, these are currently just discussions, not a partnership or even an acquisition. Whether Apple and PrismML will actually collaborate remains to be seen, and there is no concrete timeframe for a potential deployment.<\/p>\n\n\n\n

A sober technical perspective is also warranted. Fitting a model with 27 billion continuously active parameters onto a smartphone with limited RAM is only possible through very aggressive compression \u2013 and the claim that no performance is lost in the process comes from the startup itself and has not yet been independently verified. Furthermore, there is a fundamental objection: Critics of purely local approaches point out that cloud models are currently improving almost weekly. A model that runs exclusively on the device could lag behind this pace in the long run.<\/p>\n\n\n\n

A possible building block for Apple's AI course<\/h2>\n\n\n\n

The fact that Apple is even exploring external compression expertise is the truly revealing aspect of the report. It suggests that the company is well aware of a gap between what its own models can currently achieve and what current iPhone hardware could theoretically handle. Whether this will lead to a concrete collaboration or remain just one of many exploratory steps will likely only become clear once PrismML makes its model publicly available \u2013 allowing third parties to verify its bold claims. (Image: Apple)<\/p>\n\n\n\n