Apple has held meetings with PrismML about integrating the startup’s technology to run substantially larger AI models directly on iPhones. According to The Information, the technical breakthrough centers on PrismML’s ability to compress Alibaba’s open-source Qwen 3.6 language model down to run on an iPhone 17 Pro. This is a 27-billion-parameter model where all parameters remain simultaneously active, surpassing Apple’s current on-device flagship in both total capacity and operational architecture.
Apple may be looking beyond its own Apple Foundation Models framework to advance on-device AI capabilities. Apple’s current largest on-device model AFM 3 Core Advanced with 20 billion parameters, powers iOS 27 features like expressive Siri voices and improved systemwide dictation on iPhone 17 Pro and iPhone Air. But only 1 billion to 4 billion of those parameters are active at any given moment. PrismML’s approach keeps all 27 billion parameters active simultaneously, a fundamentally different optimization strategy that trades some efficiency for raw capacity.
PrismML plans to release its open-source model publicly on July 14. The company has already demonstrated production-ready on-device AI through Bonsai Studio, its image generation app, which generates 512×512 images on iPhone 17 Pro in roughly 12 seconds using the firm’s ternary Bonsai Image model. The startup compresses models that normally require server infrastructure to run locally with minimal latency and memory footprint on mobile silicon.
Denser models mean richer local capabilities without waiting for cloud responses. A 27-billion-parameter model kept fully active can handle nuanced writing tasks, code generation, and contextual reasoning that sparse 1-4 billion active-parameter models cannot. The advantage compounds for users in regions with spotty connectivity or those who prioritize offline functionality.
For Apple, the strategic angle is direct as larger models running on-device reduce dependence on Private Cloud Compute, Apple’s hybrid approach where sensitive queries are processed on Apple-owned servers rather than the cloud. Shifting more compute to the device itself lowers operational costs and reinforces Apple’s privacy-first positioning against competitors like Google and OpenAI, both of which funnel substantial processing to remote servers. It also unlocks more capable local features; PrismML’s Qwen 3.6 is capable of software development tasks, substantially more complex than typical on-device assistants.
Apple is actively exploring third-party optimization technologies, which suggests the company recognizes a gap between what its own models deliver and what on-device hardware can theoretically support. This doesn’t necessarily mean acquisition (though Apple has bought AI startups before), but it reveals Apple views external expertise in model compression as worth investigating. Whether PrismML’s dense-parameter approach becomes standard in iOS is not clear yet, but the conversation itself shows that on-device AI capacity on iPhones is still far from its ceiling.
