Rismml overcame the end-side AI bottleneck: successfully “stuffed” Ali Tongyi's thousand questions into the Apple (AAPL.US) iPhone 17 Pro, running with all parameters comparable to the cloud

The Zhitong Finance App learned that, according to people familiar with the matter, artificial intelligence startup Prismml has compressed Alibaba (BABA.US)'s open source big language model Qwen 3.6 so that it can run on Apple's (AAPL.US) iPhone 17 Pro.

According to people familiar with the matter, Apple has already held a meeting with PrisMML to discuss how to use its technology.

The model has 27 billion (27B) parameters, and the number of parameters can help determine the complexity of the data that the model can handle. In contrast, most models that run on phones have only a few billion parameters active at a time.

The largest artificial intelligence model can have trillions of parameters. Currently, it is still too large to run on mobile devices. However, according to the startup, the model PrisMML successfully runs on the iPhone, capable of performing tasks such as complex dialogue, reasoning, completely autonomous agents (agents), and software coding. The report added that the open source model will be available for download next Tuesday.

PrisMML uses a mathematical trick to reduce the size of the Qwen 3.6 model to a fraction of its original size. Often, downsizing the model causes performance to degrade, but the company claims that its approach to miniaturizing the AI model doesn't hurt its performance. According to reports, PrismML has shrunk the size of Qwen 3.6 from about 54 megabytes (GB) to less than 4 megabytes (GB).

This milestone reflects a broader trend of driving artificial intelligence to run on the device side rather than on expensive high-power servers in data centers. Companies such as Microsoft (MSFT.US), Amazon (AMZN.US), and Meta Platforms (META.US) are investing hundreds of billions of dollars to build data centers to meet their anticipated demand for artificial intelligence.

However, while Apple has largely remained on the sidelines of this data center race, it is also a staunch proponent, making sure that many of the iPhone's artificial intelligence features run on the device (not the cloud) as much as possible.

PrisMML's approach might appeal to Apple. At the company's Global Developers Conference (WWDC) in June of this year, Apple announced a complete overhaul of Siri based on the Google (GOOGL.US) Gemini model. The most advanced part of Siri is still so large that Apple needs to use Nvidia (NVDA.US) chips running in Google Cloud. As part of the new Siri announcement, Apple indicated that some of the iPhone's new AI features will run on the device side.

Apple's new end-side model has 20 billion (20B) parameters, but uses a so-called sparse architecture (sparse architecture), where only 1 billion to 4 billion (1B to 4B) parameters are active at one time. People familiar with the matter added that in PrisMML's end-side model, all 27 billion parameters are activated simultaneously.

As OpenAI's first venture capitalist, Khosla Ventures (Khosla Ventures) invested PRISMML's $16.25 million seed round earlier this year.

PrisMML is a California Institute of Technology spin-off, and its CEO is Babak Hasibe. Hasibe is a professor of electrical engineering at the university, and he and the co-founder conducted mathematical research on the startup's technology. According to the report, Caltech holds the patent behind this technology, but has exclusively licensed it to PrisMML.