The 27 billion-parameter AI 'Qwen3.6-27B' runs on the iPhone 17 Pro; Apple is reportedly in discussions with PrismML and compression technology.

AI development company PrismML claims to have successfully compressed its large-scale language model 'Qwen3.6-27B,' which has 27 billion parameters, and made it run on an iPhone 17 Pro alone. The Information also reports that Apple met with PrismML to discuss how to use PrismML's technology on iPhones.

Khosla-Backed Startup Claims Breakthrough With Largest-Ever AI Model on an iPhone — The Information
https://www.theinformation.com/articles/khosla-backed-startup-claims-breakthrough-largest-ever-ai-model-iphone

Apple Exploring Ways to Run Much Larger AI Models Directly on iPhones - MacRumors
https://www.macrumors.com/2026/07/09/apple-prismml-larger-on-device-ai-models/

Apple Explores Larger On-Device AI Models with PrismML
https://www.ithinkdiff.com/apple-prismml-on-device-ai-models/

Apple cites the advantages of its in-device-running base model, such as its ability to operate offline and its ease of keeping personal information on the device. However, fitting a high-performance AI model into a smartphone is not easy. The numerical values that the AI model uses to generate text and make decisions are called 'parameters,' and the more parameters there are, the more memory and computing power is required.

Apple's 'AFM 3 Core Advanced,' announced in June 2026, has a total of 20 billion parameters, but employs a sparse architecture that enables only 1 billion to 4 billion parameters depending on the processing task. The entire model is stored in flash memory, and only the necessary parts are moved to working memory, enabling it to operate on the iPhone.

What's so amazing about the new AI 'Apple Foundation Models'? A groundbreaking mechanism that runs the 20 billion parameter multimodal model 'AFM 3 Core Advanced' on an iPhone - GIGAZINE

PrismML has developed a method to significantly reduce the number of bits used to represent the weights of AI models, thereby lowering the required memory capacity and computational load. While typical AI models store each parameter using fine-grained numerical values such as 16 bits, PrismML's '1-bit Bonsai' represents each weight with 1 bit, and 'Ternary Bonsai' represents them with three types of values: '-1', '0', and '1'. PrismML explains that 1-bit Bonsai can reduce memory usage to up to 1/14th, processing speed to up to 8 times faster, and energy consumption to up to 1/5th compared to the 16-bit version.

A memory-efficient AI model, '1-bit Bonsai,' has appeared, boasting 8B of parameters but consuming only 1.15GB of memory, while achieving performance equivalent to or better than models consuming 14 times more memory - GIGAZINE

The Qwen3.6-27B model that PrismML claims to have run on an iPhone 17 Pro is a 27 billion parameter dense model released by Alibaba's Qwen team. Unlike models that select only some parameters, dense models use the entire model for processing. According to The Information, PrismML compressed the model, which was originally about 54GB, to less than 4GB and ran it with all 27 billion parameters enabled simultaneously.

PrismML has released not only a language model, but also an app called 'Bonsai Studio by PrismML' that generates images on an iPhone. Bonsai Studio is said to complete the image generation process entirely within the iPhone 17 Pro and can generate a 512x512 pixel image in about 12 seconds.

I tried out 'Bonsai Image 4B,' an image generation AI that runs locally on iPhones, and modified FLUX.2 Klein 4B into a 1-bit version to reduce memory usage to 1/8.3 - GIGAZINE

If a model with 27 billion parameters can run on an iPhone alone while maintaining practical speed and generation quality, it may be possible to complete processes such as complex document creation and program code generation entirely within the iPhone. For Apple, this technology could reduce the need to send Apple Intelligence processing to Private Cloud Compute, thereby lowering server operating costs while enhancing privacy protection. However, it remains unclear whether Apple will adopt PrismML technology or pursue a partnership or acquisition.

PrismML has stated that it plans to release a compressed version of Qwen3.6-27B as open source on July 14, 2026.