Apple’s Core Advanced model has 20 billion parameters stored in flash. It activates only 1 to 4 billion per prompt. The model runs on the Neural Engine at about 4 watts. It uses 2.0 to 3.4 GiB of memory including safety models. It supports 8,192 tokens of context. It is built in collaboration with Google. It powers Siri’s new voices and dictation. It is designed for tasks like summarizing and classification. It avoids basic math, code and logical reasoning. It is available on macOS 27 for Macs with M3 or later and at least 12 GB RAM.
What changed
On-device models were fully loaded in RAM and used all parameters for every prompt.
What it unlocks
Running a 20-billion-parameter model on a Mac with limited memory and battery.
- 20 billion parameters
- 1 to 4 billion active per prompt
- 8,192 token context
- 4.95 GB model file in flash
What you need to act on it
- macOS 27
- Mac with M3 or later
- at least 12 GB RAM
Sources