Apple Runs 20-Billion-Parameter AI On iPhone With Sparse Activation
Apple proved that a 20-billion-parameter model can run locally on an iPhone, shifting the AI competition from cloud-only scale to on-device capability.
Key Facts
- Apple's third-generation AI foundation model, AFM 3 Core Advanced, has 20 billion parameters and runs on the iPhone using a sparse activation architecture.
- Apple says the on-device model scored 4.24 versus 3.82 on a 5-point scale for speech synthesis, and was preferred in speech input quality and comprehension.
- Server-side models AFM 3 Cloud and AFM 3 Cloud Pro handle complex reasoning and image tasks via Private Cloud Compute, with the Pro model using Google Cloud and NVIDIA GPUs.
Reporting from 2 sources: ASCII.jp, GIGAZINE.
Apple announced the third-generation Apple Foundation Models at WWDC26, including the 20-billion-parameter AFM 3 Core Advanced that runs on-device via a sparse activation architecture. The model loads only necessary parts from NAND flash into DRAM, enabling practical speeds. Speech synthesis and input quality improved significantly. Server-side AFM 3 Cloud and AFM 3 Cloud Pro handle complex tasks.
Apple's third-generation AI foundation model, AFM 3 Core Advanced, packs 20 billion parameters but runs on the iPhone by using a sparse activation architecture. Instead of loading the entire model into DRAM, the system stores it in NAND flash and pulls only the specialized parts needed for each query. Apple says this approach lets a model larger than the device's memory capacity operate at practical speeds. The on-device model outperformed conventional systems in speech synthesis, scoring 4.24 versus 3.82 on a 5-point scale, and was preferred in speech input quality and comprehension. Server-side models AFM 3 Cloud and AFM 3 Cloud Pro handle complex reasoning and image tasks via Private Cloud Compute, with the Pro model also leveraging Google Cloud and NVIDIA GPUs.
Synthesized by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited, then editorially reviewed before publishing.