Apple Unveils Enhanced On-Device and Server Foundation Language Models at WWDC 2025

June 9, 2025

Apple has made significant strides in integrating advanced generative AI into its ecosystem with the latest updates to its foundation language models. At the 2025 Worldwide Developers Conference (WWDC), Apple introduced a new generation of these models, specifically designed to enhance the Apple Intelligence features across its platforms. The company also launched the Foundation Models framework, providing app developers direct access to the on-device foundation language model that powers Apple Intelligence.

Key Technical Updates

The latest updates focus on improving tool-use and reasoning capabilities, expanding multilingual support, and enhancing efficiency. Here’s a breakdown of the key changes:

• Improved Tool-Use and Reasoning: The new models can better understand and interact with various tools and environments, making them more versatile for tasks that require logical reasoning.
• Multimodal Inputs: They now handle both image and text inputs, expanding their utility in mixed-media applications.
• Speed and Efficiency: Optimized to run faster and use fewer resources, particularly on Apple silicon devices.
• Language Support: The models support 15 languages, making them more accessible globally.

Model Architectures

Apple developed two primary foundation models: an on-device model and a server-based model. Both are designed to meet different performance and deployment requirements.

On-Device Model

• Architecture: This model is compact, with approximately 3 billion parameters.
• Optimization: Tailored for Apple silicon, it ensures low-latency inference and minimal resource usage.
• Use Cases: Ideal for real-time applications like voice assistants, text predictions, and image recognition.

Server-Based Model

• Architecture: A mixture-of-experts model with a novel architecture designed for Private Cloud Compute.
• Optimization: Focuses on high accuracy and scalability for more complex tasks.
• Use Cases: Suitable for applications requiring deep reasoning, large-scale data processing, and high computational demands.

Training and Optimization

To achieve these advancements, Apple employed several techniques:

• Data: The models were trained on a diverse dataset that includes text, images, and tool interactions to ensure robust performance across different scenarios.
• Training Recipes: Advanced training methods were used to fine-tune the models for specific tasks and environments.
• Inference Optimization: Techniques like model pruning, quantization, and efficient inference pipelines were applied to enhance speed and reduce resource consumption.

Evaluation Results

Apple’s evaluation of these new models shows significant improvements in both performance and efficiency. Key benchmarks include:

• On-Device Model:
  • Latency: Reduced by 30% compared to the previous generation.
  • Accuracy: Improved by 15% on standard NLP tasks.
• Server-Based Model:
  • Scalability: Handles up to 10x more concurrent requests than the previous model.
  • Accuracy: Achieves state-of-the-art results on complex reasoning tasks.

Responsible AI

In line with Apple’s commitment to privacy and responsible AI, the development process integrates several principles:

• Privacy: The on-device model processes data locally, ensuring user information remains private.
• Transparency: Clear documentation and guidelines are provided for developers using the Foundation Models framework.
• Fairness: Rigorous testing is conducted to minimize biases and ensure fair performance across different user groups.

Conclusion

With these updates, Apple continues to push the boundaries of AI integration in its products. The new foundation models not only enhance the intelligent features of Apple’s ecosystem but also provide developers with powerful tools to create innovative applications. By focusing on efficiency, multilingual support, and responsible AI practices, Apple is setting a high standard for the future of generative AI.