Exploring Apple's On-Device Model: Twenty Billion Parameters in Flash
The article looks at Core Advanced, the 20-billion-parameter model Apple ships on Macs that run Apple Intelligence, and how it fits into a Mac's memory. The author reverse-engineers nothing; instead they measure the model's footprint and behavior from outside using macOS tools and public commands.
- Apple's design keeps most of the model in flash storage and swaps in a few experts per prompt, so two seconds into a long request it holds about 2.0 GiB of memory versus roughly 15 GiB for an open model of similar class.
- The tradeoff is capability: on 20 math problems it finished last of four, and developers can't pin or tune the model, while Apple's own docs list math, code and logical reasoning as tasks to avoid.
- Apple keeps adapters for its own features while closing them to developers, and runs safety models as a separate process, so it can update screening without touching the main model.
Apple's Core Advanced is the larger of its two on-device language models, shipped with macOS 27 on any Mac that runs Apple Intelligence. It has 20 billion parameters, but only 1 to 4 billion are active at a time, and Apple says it was built in collaboration with Google. It powers Siri's new voices and dictation, and any script can call it through a terminal command like `fm respond`. The author could not read Apple's weights, since the license forbids reverse engineering, so the analysis rests on what macOS exposes from outside: file reads, system logs, memory and power counters, and Apple's public commands. Tested on an M3 Max MacBook Pro with 48 GB, the findings are collected in a public repository where each claim is labeled as Apple's statement, a measurement, an inference, or an explicit assumption.
The headline numbers are the efficiency ones. Two seconds after a long request, Apple's model and its safety models held 2.0 GiB of memory, against roughly 15 GiB for an open model of the same class built conventionally. On 20 math problems, Core Advanced used about a third of the energy of the best open build tested, but it finished last of four on accuracy. The trick is architectural: the model's 4.95 GB of weights sit in flash storage, a small dense block picks a fixed set of experts per prompt rather than per token, and only 636 MiB stays resident on the Neural Engine. Apple's own research explains the reasoning, since flash-to-RAM bandwidth is too slow to swap experts token by token. The Mac's read pattern matches this: a few blocks read at a steady rhythm every 32 tokens, which the author reads as periodic re-selection of experts.
The tradeoffs are clear and the design is plainly aimed at short, frequent jobs, not general chat. Apple's developer guide lists summarizing, entity extraction, classification and creative writing as intended uses, and says to avoid basic math, code and logical reasoning. The model runs beside a separate safety process with its own two 300-million-parameter classifiers, which Apple can update independently, though those cost 0.7 GiB of the total. Developers cannot pin or tune the model, and Apple has dropped third-party adapters in OS 27 while keeping its own, so a custom adapter trained on this year's base would likely break at each update. Requests the device cannot handle go to Apple's Private Cloud Compute, and the app must choose that path explicitly.