TLDR
- Apple is in talks with Caltech spinout PrismML about using its on-device AI compression technology
- PrismML shrunk Alibaba’s 27-billion-parameter Qwen 3.6 model from 54 GB down to under 4 GB
- Unlike other compressed models, all 27 billion parameters remain active simultaneously on the iPhone 17 Pro
- Apple’s current advanced Siri features still rely on Google’s Gemini running in the cloud via Nvidia chips
- PrismML raised a $16.25M seed round backed by Khosla Ventures earlier this year
Apple (AAPL) stock was up 0.44% on Wednesday as reports emerged the company is in talks with AI startup PrismML, a spinout from the California Institute of Technology.
The discussions, first reported by The Information, center on PrismML’s ability to run a 27-billion-parameter AI model directly on an iPhone 17 Pro — without sending data to the cloud.
PrismML compressed Alibaba’s open-source Qwen 3.6 model from roughly 54 gigabytes down to less than 4 gigabytes. That’s a reduction of more than 90%.
What makes this different from other shrinking efforts is performance. Most compressed models sacrifice accuracy to fit on a device. PrismML claims it doesn’t.
The startup achieves this using ultra-dense 1-bit and ternary weight architectures, which cut memory requirements by up to 14x and run up to 8x faster than standard models.
Most on-device AI models only activate a fraction of their parameters at once — a method called sparse architecture. PrismML keeps all 27 billion parameters active at the same time.
That gives the on-device model the ability to handle complex tasks like reasoning, autonomous agents, and software coding, according to the startup.
The open-source version of the model is expected to be available for download next Tuesday.
Apple’s Cloud Dependency Problem
Right now, Apple’s most advanced Siri features — announced at WWDC in June — rely on Google’s Gemini models running on Nvidia chips inside Google Cloud data centers.
Apple does have one new on-device model with 20 billion parameters, but only 1 to 4 billion are active at any given time under its sparse setup.
PrismML’s full-parameter approach would be a step up. It would let Apple run heavier AI workloads entirely on-device, cutting cloud costs and keeping user data off external servers.
Who Is PrismML?
PrismML was founded by Babak Hassibi, a professor of electrical engineering at Caltech. The mathematical research behind the technology was developed at the university.
Caltech holds the patents on the core technology but licenses them exclusively to PrismML.
The startup raised a $16.25 million seed round earlier this year, backed by Khosla Ventures — the same firm that made the first VC bet on OpenAI.
Apple and PrismML have not publicly commented on the acquisition talks.
While rivals like Microsoft, Amazon, and Meta are spending hundreds of billions building out data center infrastructure, Apple’s strategy has consistently leaned toward keeping AI on the device.
A deal with PrismML would push that strategy further than anything Apple has done to date.
4th of July Flash Sale – 50% OFF!
Celebrate Independence Day by investing in your future. For a limited time, get 50% OFF a Knockout Stocks membership and unlock our latest high-conviction stock picks, powered by our proprietary KO Score algorithm.
You'll also get access to our long-term investment ideas and shorter-term trade opportunities, helping you identify potential opportunities before the crowd.
Sign up to Knockout Stocks today and get 50% OFF to unlock the full list of premium stock picks.
Use coupon code SPECIAL50 for your exclusive discount.
Offer ends soon. Don't miss out!
Get 3 Free Stock Ebooks
Discover top-performing stocks in AI, Crypto, and Technology with expert analysis.
- Top 10 AI Stocks - Leading AI companies
- Top 10 Crypto Stocks - Blockchain leaders
- Top 10 Tech Stocks - Tech giants
