Apple eyes larger on-device AI models for iPhone with PrismML prototype

Apple has reportedly held discussions with startup PrismML about using its technology to run much larger on-device AI models for iPhone, a shift that could expand Apple Intelligence features and reduce dependence on cloud compute. According to a report from The Information cited by MacRumors, PrismML demonstrated a 27-billion-parameter Qwen 3.6 model running entirely on an iPhone 17 Pro, surpassing Apple's current on-device AFM 3 Core Advanced model at 20 billion parameters.

What happened

Apple has held meetings with PrismML about ways to deploy the startup's technology to run larger AI models directly on iPhones. The report states that PrismML managed to shrink down Alibaba's open-source Qwen 3.6 to run on an iPhone 17 Pro. The model has 27 billion parameters, larger than Apple's on-device AFM 3 Core Advanced model with 20 billion parameters, which powers iOS 27 enhancements such as Siri AI's more expressive voices and improved systemwide dictation on iPhone 17 Pro and iPhone Air models.

A key difference: unlike AFM 3 Core Advanced, all of Qwen 3.6's parameters can be active at the same time. The report notes that Apple's model uses a sparse architecture where only 1 to 4 billion parameters are active at a time, while PrismML's on-device model keeps all 27 billion parameters active simultaneously.

Why AI builders should care

For teams building AI products or workflows, this development highlights a growing push toward denser on-device inference. Larger on-device AI models for iPhone could enable more capable local inference, expanding features like Siri voices and dictation without routing data to Apple's Private Cloud Compute servers. This could reduce operational costs and improve user privacy, a key consideration for consumer-facing AI apps.

Hardware-software teams should note that the PrismML approach uses a dense model architecture, which contrasts with Apple's sparse approach. This suggests that model compression and optimization techniques (like those used by PrismML) may become more important for delivering high-capability on-device AI without sacrificing performance.

Practical implications

If feasible, running larger models on-device could shift some workloads away from cloud compute, affecting both costs and data privacy posture. The 27B model is described as fully active (dense) on-device, while the 20B AFM 3 Core Advanced model uses sparsity (1-4B active at a time). This means PrismML's approach may offer more consistent reasoning capability per inference, but it also raises questions about memory bandwidth, power consumption, and thermal management on mobile devices.

For iOS 27 AI features, the report ties Siri voices and dictation to the on-device model, but the larger parameter count could enable more complex tasks like real-time language understanding, multimodal processing, or personalized recommendations without a network round trip.

Caveats

This is an early-stage report. The details come from a single source citing discussions and a demonstration, not an official Apple announcement. Feasibility, performance, and official support remain unconfirmed. The demonstration was on an iPhone 17 Pro, which may have specific hardware (e.g., A19 chip, improved neural engine) not available in older models. Battery life, inference speed, and real-world usability are not addressed in the source. Builders should treat this as a signal of direction rather than a shipping capability.

FAQs

Can the iPhone run a 27-billion-parameter AI model locally?

The MacRumors article reports that PrismML demonstrated a 27-billion-parameter Qwen 3.6 model running on an iPhone 17 Pro, but this remains a reported discussion and demo, not an official Apple release. No feasibility or performance guarantees are provided in the source.

What is the difference between PrismML on-device models and AFM 3 Core Advanced?

According to the report, PrismML keeps all 27 billion parameters active on-device (dense architecture), while Apple's AFM 3 Core Advanced uses a sparse architecture where only 1 to 4 billion parameters are active at a time.

How could larger on-device AI models affect privacy and reliance on cloud compute?

The article suggests that running larger models on-device could reduce reliance on Apple's Private Cloud Compute and enhance user privacy, but this claim is based on reported discussions and is not independently confirmed.

Which iPhone models support running larger on-device AI models?

The discussed demonstration references an iPhone 17 Pro. No official compatibility list or broader device support is provided in the cited material.