Julien Simon investigates Apple's Core Advanced model

Julien Simon, a former chief evangelist at Hugging Face, published a detailed analysis of Apple's Core Advanced language model on October 25 in his newsletter The AI Realist. He reconstructed a model card for Core Advanced, as Apple had not officially released one, using a MacBook Pro M3 Max with 48 GB of memory.

Core Advanced's resource usage

Simon's measurements showed that a 20 billion parameter model like Core Advanced consumes as much memory and energy as a 4 billion parameter model on Apple's Macs. This is due to the way the model is structured and managed within the hardware.

Core Advanced's capabilities and deployment

Apple introduced Core Advanced on June 8, 2026, during the WWDC, describing it as the largest of its two embedded models in a family of five "custom-designed in collaboration with Google." It has 20 billion parameters but only 1 to 4 billion are active at any given time, used for features like Siri's new expressive voices and dictation.

Core Advanced's hardware requirements

Apple reserves Core Advanced for Macs with at least 12 GB of unified memory, such as the Mac M3 or newer models. On mobile devices, it powers the iPhone Air, iPhone 17 Pro, 17 Pro Max, and the iPad M4 or newer models with 12 GB of memory.

Core Advanced's architecture

Core Advanced is based on a mixture of experts (MoE) architecture, where each layer is divided into specialized networks. Unlike traditional MoE models, Apple keeps the full model in storage and selects a subset of experts dynamically during a query, reducing memory usage.

Core Advanced's performance on hardware

The model runs on the Neural Engine, a dedicated AI chip in the M3 Max, consuming about 4 watts of power. The graphics processing unit (GPU) remains idle during these operations, indicating that the Neural Engine is the primary component handling AI tasks.

Core Advanced's memory and compression

Simon estimated that the model's weights are compressed to about 2 bits each. Two security models with 300 million parameters each run separately and filter requests. During a long query, they used up to 3.4 GB of memory, compared to 15 GB used by a similar open-source model.

Core Advanced's performance benchmarks

In tests, Core Advanced solved only 13 out of 20 problems, compared to 20 for the best open model, Qwen3 4B. It performed better on general knowledge questions but lagged in logical reasoning and basic calculations, aligning with Apple's documentation that advises against using it for such tasks.

Comparison with open-source models

Apple's model is more efficient in terms of energy use but less accurate in some tasks compared to open-source models like Gemma 4 26B A4B and Qwen3 4B. The difference in performance may be partly due to the compression techniques used, with Apple's 2 bits per weight being close to the most advanced methods.

Apple's technical report delay

Apple had promised to publish a technical report on Core Advanced by summer, but as of October 25, it had not been released. This delay may affect developers and researchers who are looking for more detailed information about the model's design and capabilities.