Apple’s broader aim appears to close the AI capability gap without the heavy cloud investments of rivals, and the company may pursue acquisitions to move workloads onto iPhones.

Unlike sparse on-device models, PrismML’s full-parameter approach keeps all 27 billion parameters active, enabling advanced offline reasoning and autonomous tasks.

The on-device approach, if adopted, could reduce reliance on cloud processing, lower Siri cloud infrastructure costs, and keep user data on the device.

PrismML’s approach leverages ultra-dense 1-bit and ternary weight architectures, delivering up to 14x memory reduction and up to 8x faster performance.

Apple may be pursuing a clearer path for on-device AI after PrismML reportedly compressed the 27‑billion‑parameter Qwen 3.6 model from about 54GB to under 4GB, enabling on-device chat, reasoning, autonomous agents, and code generation.

PrismML plans to release an open-source version of the model next Tuesday.

If negotiations succeed, this would mark a significant acceleration of Apple’s device-centric AI strategy amid a wider industry shift toward centralized data centers by competitors.

Industry observers see both support for on-device AI and the challenge of hybrid architectures that blend on-device work with cloud reasoning to capitalize on rapid cloud-model advances.

PrismML is led by Babak Hassibi and backed by California links and Khosla Ventures, with projections that broader on-device AI adoption could arrive within about three years.

The reporting places PrismML developments in the broader context of WWDC 2026 and iOS 27 features, including enhanced dictation and Home app integrations, shaping Apple’s AI landscape.

The technology promises on-device AI capability comparable to larger server-based models, including tasks like software development.

Apple’s wider AI moves include collaboration with Google on Siri AI and strategic acquisitions such as Q.ai, indicating a push to expand on-device AI capabilities.