27B Model Runs on iPhone! Apple's AI Compression Breakthrough
Apple is making moves to supercharge its on-device AI. According to tech media outlet The Information, the company is in talks with PrismML, a Caltech spin-off, to explore running larger AI models directly on iPhones. If successful, this could mean your next iPhone handles complex AI tasks without needing the cloud.
The Secret Sauce: 1-Bit Compression
PrismML's breakthrough is a native 1-bit model compression technology. Traditional quantization reduces precision but still uses multi-bit weights. PrismML goes further: it represents weights using only -1 and +1, combined with group scaling factors. This cuts model size to about one-fourteenth of the full-precision version and reduces memory usage by over 90%.
What's remarkable is that this approach doesn't rely on a "high-precision escape route"—a common trick where key layers retain high precision to compensate for performance loss. Despite using 1-bit weights across the board, PrismML claims its models maintain accuracy close to FP16 models. Inference speed can increase up to eight times, and energy consumption drops by 75% to 80%. That's a game-changer for mobile devices.
Real-World Proof: 27B Model on iPhone 17 Pro
The most compelling evidence comes from a live demo. PrismML compressed Alibaba's open-source Qwen 3.6 model—a 27-billion-parameter giant—and ran it entirely on an iPhone 17 Pro. Under normal circumstances, such a large model would be impossible to run smoothly on a smartphone. But after compression, it not only ran but maintained near-original inference quality.
Apple is reportedly impressed by this capability and sees it as a way to enhance the reasoning performance of local AI models. For Apple, stronger on-device AI means a more competitive Apple Intelligence ecosystem. Currently, iPhone AI models are limited by memory and power constraints. With PrismML's tech, iPhones could run larger models without hardware upgrades, enabling complex multi-turn conversations, image understanding, and intelligent task orchestration.
What This Means for the Future
If Apple integrates PrismML's compression, the balance between on-device and cloud AI could shift dramatically. Flagship-level large models running on smartphones would reduce reliance on cloud servers, improving privacy and responsiveness. It's a bold step toward making advanced AI truly portable.

Key Points
- Apple is in talks with PrismML to bring large AI models to iPhones.
- PrismML's 1-bit compression reduces model size to 1/14 and memory usage by 90%+.
- Inference speed increases up to 8x, energy consumption drops 75-80%.
- A 27B-parameter Qwen model already runs on iPhone 17 Pro.
- This could enable complex AI tasks on-device, reducing cloud dependence.