PrismML Compresses 27B-Parameter AI Model to Run on iPhone 17 Pro
If PrismML's compression proves practical, Apple could run large dense AI models entirely on-device, reducing server costs and improving privacy without relying on cloud processing.
Reporting from 1 source: GIGAZINE.
AI developer PrismML claims to have compressed Alibaba's 27-billion-parameter Qwen3.6-27B model to under 4GB and run it on an iPhone 17 Pro with all parameters active. Apple has held a meeting with PrismML to discuss using the compression technology. PrismML's 1-bit Bonsai method reduces memory and computation. The compressed model is planned for open source release on July 14, 2026.
PrismML's 1-bit Bonsai method represents each weight with a single bit, cutting memory usage by up to 1/14 compared to standard 16-bit models. The company claims it compressed the approximately 54 GB Qwen3.6-27B to under 4 GB while keeping all 27 billion parameters active. Apple's current on-device model, AFM 3 Core Advanced, uses a sparse architecture that activates only 1 to 4 billion parameters at a time. Apple has reportedly met with PrismML to discuss integrating the technology into iPhones. PrismML plans to release the compressed model as open source on July 14.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.