Apple Negotiates with PrismML to Compress AI Models Like Qwen 3.6 for iPhone

Apple is in talks with startup PrismML about the possible use of its technology for compressing large language models, which will allow running powerful AI models directly on iPhone without connecting to the cloud — The Information reported this citing its sources.

What PrismML has already managed to compress

Talks began after PrismML achieved a technological breakthrough: the startup successfully compressed the open Qwen 3.6 model from Alibaba for local execution on iPhone 17 Pro.

• The Qwen 3.6 model from Alibaba contains 27 billion parameters
• PrismML compressed it without loss of functionality and intellectual capabilities to run on iPhone 17 Pro
• The technology allows powerful AI models to be used without connecting to the cloud
• The Information reported on talks between Apple and PrismML citing sources

How PrismML solves the overheating problem

To fit a model with such a number of parameters on a smartphone usually requires a "sparse architecture" — the device activates only part of the AI "brain" at any given moment to avoid overheating. According to available data, PrismML completely changed this approach to model compression.

What this means

If the deal goes through, Apple will gain the ability to run truly large AI models directly on iPhone without the cloud. In recent years, the company has bet on private AI that works without sending data outside — a partnership with PrismML fits this strategy and could be a significant step toward more powerful and faster AI features on the device.

Frequently asked questions

What is PrismML?

It is a startup that has developed technology for significant compression of large language models without loss of their functionality and intellectual capabilities — according to The Information, Apple is in talks with it.

What model has PrismML already compressed?

The open Qwen 3.6 model from Alibaba with 27 billion parameters — it managed to compress it for local execution on iPhone 17 Pro.

On what iPhone was the compressed model already run?

The Information's report mentions iPhone 17 Pro — it is on this device that PrismML achieved local execution of the compressed Qwen 3.6 version.