Reproduce “Apple Intelligence Foundation Language Models: Tech Report 2025” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ethan Li, Anders Boesen Lindbo Larsen, Chen Zhang, Xiyou Zhou, Jun Qin, Dian Ang Yap, Narendran Raghavan, Xuankai Chang and 389 more
The authors describe two language models for Apple devices and Private Cloud Compute, tailoring architecture and compression to each deployment. The on-device model shares key-value cache and is optimized to 2 bits per weight; the server model uses parallel-track mixture-of-experts. MMLU falls from 67.8 to 64.4 for the on-device model after optimization; these scores do not establish unchanged performance elsewhere.
Key insights
- A dual-model strategy effectively addresses the trade-off between on-device efficiency/privacy and server-side capability, leveraging Apple's hardware-software co-design.
- Aggressive quantization (2-bit QAT) and compression techniques (ASTC with LoRA) can significantly reduce model footprint and improve inference speed on consumer hardware with minimal quality degradation.
- A "Swift-centric Foundation Models framework" with guided generation and constrained tool calling simplifies AI integration for developers, fostering an ecosystem of intelligent applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.