Apple has introduced two new multilingual foundation language models that power its Intelligence features across various devices and services. The models are designed to optimize performance on Apple silicon and the Private Cloud Compute platform.
The first model, built for on-device use, is approximately 3 billion parameters and utilizes novel architectural innovations such as KV-cache sharing and 2-bit quantization-aware training. This allows it to deliver high-quality results while maintaining competitive costs.
In contrast, the second model is a scalable server-based architecture that combines track parallelism, mixture-of-experts sparse computation, and interleaved global-local attention. It was trained on large-scale datasets sourced from responsible web crawling, licensed corpora, and synthetic data, before undergoing fine-tuning with supervised learning and reinforcement learning.
Both models outperformed open baselines in public benchmarks and human evaluations. The introduction of a new Swift-centric Foundation Models framework allows developers to easily integrate guided generation, constrained tool calling, and LoRA adapter fine-tuning into their applications with just a few lines of code.
The advancements in Apple’s language models are grounded in its commitment to responsible AI practices, including content filtering, locale-specific evaluation, and data protection measures such as Private Cloud Compute.
Source: https://machinelearning.apple.com/research/apple-foundation-models-tech-report-2025