Key Highlights
- Cupertino-based tech giant reportedly negotiating with Caltech-born startup PrismML regarding innovative on-device AI compression capabilities
- The startup successfully reduced Alibaba’s Qwen 3.6 model (27 billion parameters) from approximately 54 GB to under 4 GB
- The compression technique maintains full activation of all 27 billion parameters simultaneously on iPhone 17 Pro hardware
- Current premium Siri functionality depends on Google’s Gemini operating through cloud infrastructure powered by Nvidia processors
- Khosla Ventures led PrismML’s $16.25 million seed funding round completed in early 2024
Apple (AAPL) stock climbed 0.44% during Wednesday trading following reports that the iPhone maker is engaged in discussions with PrismML, an artificial intelligence startup that emerged from the California Institute of Technology.
Apple Inc., AAPL
According to initial reporting from The Information, the negotiations focus on PrismML’s proprietary technology that enables a 27-billion-parameter artificial intelligence model to operate entirely on an iPhone 17 Pro device — eliminating the need for cloud-based processing.
PrismML managed to compress the open-source Qwen 3.6 model from [[LINK_START_2]]Alibaba’s[[LINK_END_2]] portfolio, shrinking it from approximately 54 gigabytes to below 4 gigabytes. This represents a compression ratio exceeding 90%.
The distinguishing factor separating this approach from alternative compression methods is sustained performance. While most miniaturized models trade accuracy for reduced size, PrismML maintains it has solved this tradeoff.
The company accomplishes this through ultra-compact 1-bit and ternary weight structures, slashing memory demands by as much as 14 times while delivering speeds up to 8 times faster than conventional architectures.
Typical on-device artificial intelligence implementations activate only a portion of their parameters simultaneously — an approach known as sparse architecture. PrismML’s innovation maintains full activation of all 27 billion parameters concurrently.
This capability enables the locally-running model to execute sophisticated operations including logical reasoning, autonomous agent functionality, and software development tasks, per the startup’s claims.
The open-source iteration of this model is scheduled for public release next Tuesday.
The Cloud Infrastructure Challenge
Presently, Apple’s most sophisticated Siri capabilities — unveiled during the June WWDC keynote — depend on Google’s Gemini models executing on [[LINK_START_3]]Nvidia[[LINK_END_3]] silicon within Google Cloud infrastructure.
Apple has developed one proprietary on-device model featuring 20 billion parameters, though its sparse architecture means only 1 to 4 billion parameters operate actively during any single moment.
PrismML’s full-parameter methodology would represent a significant advancement. It would enable Apple to execute more demanding AI operations completely on-device, reducing cloud infrastructure expenses and maintaining user information within local hardware.
Background on PrismML
Babak Hassibi, an electrical engineering professor at Caltech, founded PrismML. The underlying mathematical research supporting this technology originated from university-based investigation.
The California Institute of Technology maintains ownership of the fundamental patents while granting PrismML exclusive licensing rights.
Khosla Ventures — the investment firm that provided OpenAI’s initial venture capital — led the startup’s $16.25 million seed funding round completed earlier this year.
Neither Apple nor PrismML has issued public statements regarding the reported acquisition negotiations.
As competitors including Microsoft, Amazon, and Meta commit hundreds of billions toward data center expansion, Apple has consistently pursued a device-centric artificial intelligence philosophy.
Securing a partnership with PrismML would represent the most aggressive execution of that strategic approach in Apple’s history.
The post Apple (AAPL) Stock: Tech Giant Explores Breakthrough On-Device AI Compression Technology appeared first on Blockonomi.