Apple Intelligence Unpacked: New Tech Report Reveals Deep Insights into AI Model Training

Apple has released a comprehensive technical report, “Apple Intelligence Foundation Language Models – Tech Report 2025,” offering an unprecedented look into the training, optimization, and evaluation of its new on-device and cloud-based foundation models. This report follows the initial announcement of Apple Intelligence at WWDC25, providing crucial details for developers and enthusiasts alike.

Key revelations include the innovative architecture of Apple’s AI. The on-device model, boasting around 3 billion parameters, is cleverly split into two blocks. This design reduces memory requirements for caching by 37.5% and significantly cuts the time to output the first token, all while preserving performance. This strategic division underscores Apple’s commitment to efficient local AI processing, even on memory-constrained devices.

For its server-side operations, Apple developed a custom architecture called Parallel-Track Mixture-of-Experts (PT-MoE) for its Private Cloud Compute platform. This sophisticated design breaks down the massive AI model into smaller, specialized subnetworks that activate only when relevant to a specific task. By integrating a new Parallel Track Transformer and MoE layers, the cloud model achieves faster, more accurate responses by processing tokens independently across multiple tracks, avoiding system-wide bottlenecks.

Addressing a significant limitation, Apple has dramatically improved multilingual support. The report details a 275% increase in multilingual data used during training, now comprising 30% of the total dataset. The tokenizer’s vocabulary has also expanded by 50%, from 100K to 150K tokens. These changes have led to substantial performance gains in non-English benchmarks, ensuring features like Writing Tools are more reliable across supported languages.

Regarding data sourcing, the report confirms that the largest portion of training data came from publicly available web content crawled by Applebot, which respects robots.txt exclusions. Additionally, Apple utilized licensed data from publishers, synthetically generated data for tasks like math and code, and over 10 billion image–caption pairs for visual understanding. This detailed disclosure highlights Apple’s unique, privacy-conscious approach to AI development, distinguishing its efforts in the competitive landscape.

阅读中文版 (Read Chinese Version)

Disclaimer: This content is aggregated from public sources online. Please verify information independently. If you believe your rights have been infringed, contact us for removal.