IPhone 17 Pro Could Run 27-Billion-Parameter Models, Apple Explores

Apple’s meetings with PrismML, a startup that has squeezed a 27-billion-parameter AI model onto an iPhone 17 Pro, reveal something the company has not said publicly: the models powering Apple Intelligence today are still nowhere near what Apple wants them to be. The gap between what fits on a phone and what actually impresses users has been the defining tension in Apple’s AI strategy, and a compression technique that shrinks a 54GB model to 4GB without degrading performance is exactly the kind of unlock that changes the math.

PrismML’s approach targets that gap directly. The startup demonstrated it could take Qwen’s 3.6 model, normally 54GB with 27 billion parameters, and compress it to 4GB while preserving response quality and speed. For context, most models designed to run locally on an iPhone carry only a few billion parameters, which limits the complexity of tasks they can handle. Twenty-seven billion parameters puts a model in a category that can manage multi-step reasoning and autonomous agent behavior, things that current on-device models struggle with.

Why Model Size Has Always Been Apple’s Ceiling

Apple’s commitment to on-device processing is not new. The company has built its privacy argument around keeping user data local since the introduction of the Secure Enclave, and that philosophy carried directly into Apple Intelligence when it launched. The problem is that the most capable large language models require storage and memory that no phone can currently provide at full size, which has forced Apple to either use smaller, less capable models on-device or route sensitive queries through its Private Cloud Compute infrastructure for heavier tasks.

Apple’s on-device AI framework already accepts image inputs alongside text, a sign that the company is expanding what it asks on-device models to do. Asking more of these models while keeping them small is a compounding problem. Better compression is one of the few levers that can resolve both constraints at once without requiring a hardware breakthrough.

What the PrismML Meetings Actually Signal

According to a report from The Information, Apple representatives have already held meetings with PrismML to discuss potential partnerships. No deal has been announced, and no timeline exists for when any resulting technology might reach users or appear in Siri. Apple holds exploratory meetings with many companies that never result in a product, so the meetings alone prove nothing about what ships.

What the meetings do reveal is where Apple’s engineering priorities are pointed. The company is not waiting for chip performance to catch up to model size. It is actively looking for software-side compression that can bridge the gap now. That framing matters because it suggests Apple sees the current state of on-device AI as a problem to be solved in the near term, not a limitation to be managed indefinitely.

What This Means for the iPhone in Your Pocket

If you are running Apple Intelligence on an iPhone 17e or an older supported device, the practical ceiling on what Siri can do locally has not changed yet. Any technology emerging from these discussions would take time to integrate, test, and ship. If Apple Intelligence is not showing up on your device at all, compression advances will not change your situation until Apple also expands hardware support.

For users already on supported hardware, the more relevant question is whether Apple can eventually close the quality gap between its on-device models and the cloud-based tools they compete with. A 27-billion-parameter model running locally, at full quality and acceptable speed, would represent a meaningful shift in what a phone can do without an internet connection. Whether PrismML’s compression holds up at that scale, across Apple’s specific chip architecture and use cases, is the question that any partnership would need to answer before any of this reaches a software update.