Apple's AI strategy is markedly different from its competitors — instead of chasing the most powerful general-purpose models, it focuses on on-device deployment and privacy protection.
On-device model: 3B parameters
At WWDC 2025, Apple unveiled the Apple Foundation Models (AFM). The on-device version has approximately 3 billion parameters and is specifically optimized for Apple Silicon.
Key technologies include:
- KV-cache sharing: The model is split into two blocks (depth ratio 5:3). The later block directly reuses the KV cache from the earlier block, reducing memory usage by 37.5%.
- 2-bit quantization-aware training: Weights are compressed to 2 bits, the embedding layer to 4 bits, and the KV cache to 8 bits.
- Low-rank adapters: These compensate for the accuracy loss caused by quantization.
The result: a 3B-parameter model compressed to run smoothly on an iPhone while maintaining sufficient capability.
Server-side model: PT-MoE
Complex tasks that cannot be handled on-device are sent to the server, but Apple uses its own Private Cloud Compute — running on Apple's custom chips in the cloud, with data never leaving Apple's secure enclave.
The server-side architecture is called Parallel-Track MoE. It consists of multiple smaller Transformer "tracks" that process tokens in parallel, synchronizing only at the input and output boundaries. This reduces synchronization overhead by 87.5%.
Privacy first
Apple's design priorities are clear:
- Process on-device whenever possible.
- Use Private Cloud Compute for tasks that must go to the cloud.
- Support 15 languages and understand text and image inputs.
This stands in stark contrast to the "everything to the cloud" approach of OpenAI, Anthropic, and Google. Apple is betting that users are willing to sacrifice some capability for privacy.
There are also benefits for developers — Apple has opened up the Foundation Models framework, allowing third-party apps to directly call the on-device model. However, its capabilities still lag behind cloud-based models like GPT-4, making it more suitable for lightweight text processing and understanding tasks.
Sources: CocoLoop, Apple Machine Learning Research