.svg){kind=spacer}
Details
- Apple released a technical report detailing its new on-device and server foundation models for Apple Intelligence.
- The on-device model, with around 3 billion parameters, and a server-based PT-MoE transformer power key Apple Intelligence features.
- Innovations include KV-cache sharing and 2-bit quantization for on-device inference, and a Parallel-Track Mixture-of-Experts architecture for the server model.
- These models were trained on multilingual and multimodal data, support 15 languages, process images, and execute tool calls, outperforming comparable open baselines.
- A new Foundation Models framework allows third-party developers to integrate these AI capabilities using Swift.
Impact
Apple’s latest models showcase technology advances in efficient, privacy-focused AI, with on-device capabilities supporting offline use and a scalable, cost-effective server model. Opening the framework to developers could spur rapid integration of these features across Apple’s ecosystem. This strengthens Apple’s market position by combining private AI services with developer access and hardware efficiency.
