What problem does Apple's AFM 3 architecture solve?

The new AFM 3 architecture resolves the long-standing issue of limited DRAM capacity for AI models on devices. Previously, large models couldn't fit into RAM, restricting their capabilities. Now, weights are stored in NAND flash, while only the necessary model segments are loaded into DRAM during processing.

How does the routing mechanism work in AFM 3 Core Advanced?

In AFM 3 Core Advanced, routing occurs once per query rather than per token. The model identifies required experts (model segments) for the query, loads them into DRAM, and processes all tokens using these experts. This reduces memory load and accelerates inference.

What advantages does AFM 3 offer for enterprises?

Businesses gain the ability to deploy powerful AI agents locally without cloud dependency, critical for regulated industries requiring data control. Complex tasks can automatically shift to cloud-based AFM 3 Cloud Pro models running on Nvidia GPUs in Google Cloud.

← All news

Artificial intelligence

Apple Overcomes Memory Limits for On-Device AI Agents

June 10, 2026

Photo: images.ctfassets.net

Quick answer

Apple introduced AFM 3, an architecture enabling 20B-parameter AI models to run on devices by storing weights in NAND flash instead of DRAM.

Apple has unveiled a groundbreaking solution for on-device AI, overcoming a critical limitation—insufficient RAM capacity. At WWDC26, the company announced the AFM 3 family of models, developed in collaboration with Google. The lineup includes two local models and three cloud-based ones, all operating within the Private Cloud Compute infrastructure. The standout is AFM 3 Core Advanced—a 20-billion-parameter model that stores weights in NAND flash instead of DRAM.

Traditional approaches required entire models to reside in RAM, capping device-side model sizes. Apple’s new architecture flips this paradigm: weights are stored in flash memory, while only the necessary model segments are loaded into DRAM for processing. This innovation is powered by the Instruction-Following Pruning (IFP) mechanism, which identifies required experts at the query stage rather than per token.

The architecture dynamically scales active parameters from 1 to 4 billion based on task complexity. Simple operations use minimal parameters, while complex tasks tap into up to 4 billion from the total 20-billion pool. However, Apple has not disclosed full details: energy efficiency, thermal performance, and automatic cloud-switching criteria remain undisclosed, with a technical report promised for this summer.

For businesses, this unlocks new opportunities for AI agent deployment. Companies can now run powerful local models, eliminating cloud dependency—a critical advantage for regulated industries. Complex tasks may automatically route to cloud-based AFM 3 Cloud Pro models running on Nvidia GPUs in Google Cloud. Yet, the lack of clear switching criteria between local and cloud models complicates compliance for organizations requiring documented inference execution locations.

Common questions

What problem does Apple's AFM 3 architecture solve?: The new AFM 3 architecture resolves the long-standing issue of limited DRAM capacity for AI models on devices. Previously, large models couldn't fit into RAM, restricting their capabilities. Now, weights are stored in NAND flash, while only the necessary model segments are loaded into DRAM during processing.
How does the routing mechanism work in AFM 3 Core Advanced?: In AFM 3 Core Advanced, routing occurs once per query rather than per token. The model identifies required experts (model segments) for the query, loads them into DRAM, and processes all tokens using these experts. This reduces memory load and accelerates inference.
What advantages does AFM 3 offer for enterprises?: Businesses gain the ability to deploy powerful AI agents locally without cloud dependency, critical for regulated industries requiring data control. Complex tasks can automatically shift to cloud-based AFM 3 Cloud Pro models running on Nvidia GPUs in Google Cloud.

Dzen feed: /feed/dzen.xml · RSS: /feed.xml