rismML breaks through the edge AI bottleneck: Successfully integrates Alibaba Tongyi Qianwen into Apple iPhone 17 Pro, achieving full parameter operation comparable to the cloud

According to informed sources, AI startup PrismML has compressed Alibaba's (BABA.US) open-source large language model Qwen 3.6 to run on Apple's (AAPL.US) iPhone 17 Pro.

The sources indicated that Apple has held meetings with PrismML to discuss how to utilize its technology.

The model has 27 billion (27B) parameters, and the number of parameters helps determine the complexity of the data the model can handle. In comparison, most models running on mobile phones only have tens of billions of parameters activated at a time.

The largest AI models can have trillions of parameters, and their size remains too large to run on mobile devices. However, according to the startup, the model that PrismML successfully runs on the iPhone can perform complex dialogues, reasoning, fully autonomous agents, and software code writing tasks. The report adds that the open-source model will be available for download next Tuesday.

PrismML uses a mathematical technique to reduce the size of the Qwen 3.6 model to a small fraction of its original size. Typically, shrinking a model leads to a decrease in performance, but the company claims that its method of miniaturizing AI models does not compromise performance. Reports indicate that PrismML has compressed the size of Qwen 3.6 from about 54 megabytes (MB) to less than 4 megabytes (MB).

This milestone reflects a broader trend toward running AI on-device rather than on expensive high-power servers in data centers. Companies like 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 AI.

However, Apple has largely remained a bystander in this data center race, while also being a staunch advocate for ensuring that many AI features on the iPhone run as much as possible on-device (rather than in the cloud).

PrismML's approach may appeal to Apple. At the company's Worldwide Developers Conference (WWDC) in June, Apple announced a complete overhaul of Siri based on Google's (GOOGL.US) Gemini model. The most advanced parts of Siri remain too large, requiring Apple to use NVIDIA (NVDA.US) chips running in Google Cloud. As part of the new Siri release content, Apple noted that some of the new AI features on the iPhone will run on-device.

Apple's new edge model has 20 billion (20B) parameters but uses a so-called sparse architecture, where only 1 billion to 4 billion (1B to 4B) parameters are activated at a time. Informed sources added that in PrismML's edge model, all 27 billion parameters are activated simultaneously As OpenAI's first venture capitalist, Khosla Ventures invested $16.25 million in PrismML's seed round financing earlier this year.

PrismML is a spin-off from the California Institute of Technology, and its CEO is Babak Hassibi. Hassibi is a professor of electrical engineering at the school, and he conducted the mathematical research used in the technology of this startup with the co-founders. Reports indicate that the California Institute of Technology holds the patents behind the technology but has exclusively licensed them to PrismML