Apple is currently evaluating PrismML, a California-based startup that utilizes technology to run AI models on iPhones. According to PrismML CEO Babak Hassibi, the technology could reduce memory consumption by up to 15 times, potentially enabling 27-billion parameter models to run on an iPhone 15 or newer.
The goal here is simple: move the intelligence from the cloud to the silicon in your pocket. Large Language Models (LLMs) are memory hogs. To run a sophisticated model, you typically need a massive amount of Unified Memory—something the iPhone’s SoC (System on Chip) simply doesn’t have in the quantities required for frontier-class models.
Enter PrismML. By slashing the memory footprint of Alibaba’s open-source Qwen model from 54 GB down to less than 4 GB, Hassibi is claiming a breakthrough that fundamentally changes the storage efficiency of the model’s weights. If Apple integrates this, Siri could become faster and more private by performing more AI processing on the device itself.
The Engineering Math: From 54 GB to 4 GB
To understand why this matters, we have to look at LLM parameter scaling. Usually, a model’s size is a direct function of its parameters and the precision (bit-depth) of those parameters. Running a 27B parameter model typically requires hardware that can handle tens of gigabytes of data just to load the model into memory before a single token is even processed.
PrismML is claiming a 15x reduction in memory usage. If Hassibi’s claims hold during Apple’s current evaluation phase, we are looking at a shift where the NPU (Neural Processing Unit) can handle complex reasoning. Apple and other companies are measuring the speed, energy efficiency, and performance of the models on devices.
- Current Bottleneck: High memory requirements for LLM inference.
- The PrismML Solution: Compressed versions of AI models.
- Hardware Target: iPhone 15 or newer.
Privacy as a Product Feature
Apple’s focus on “On-Device Processing” is intended to make Siri faster and more private because more AI processing is performed on the device itself, rather than sending requests to the cloud.
If a 27B parameter model can live on an iPhone, more AI processing could be performed locally. We’re talking about a Siri that could potentially analyze information on the device itself, rather than sending requests to the cloud.
The Market Reality Check
Apple’s stock dipped 0.9 percent following the news. Why? Because “evaluating” is not the same as “shipping.” Hassibi explicitly stated that conversations are in a very early stage and it is unclear what they will lead to.
Furthermore, analysts say that AI will still require many chips.
| Metric | Standard Qwen Model | PrismML Compressed Version |
|---|---|---|
| Memory Footprint | ~54 GB | < 4 GB |
| Deployment Target | Server/Cloud GPU | iPhone 15+ (On-Device) |
| Parameter Count | 27 Billion | 27 Billion |
| Memory Reduction | Baseline | Up to 15x |
The Verdict for Power Users
The technical hurdle remains: can PrismML maintain the logic and nuance of a 27B model at 4GB?
