Apple’s largest on-device model holds about 2.0 GiB of memory on a Mac, while a comparable open model held roughly 15 GiB under the same test. Julien Simon, writing in the AI Realist newsletter, ran both on an M3 Max laptop with 48 GB of memory, and the gap traces back to one design decision with a price attached.
The model is Core Advanced, part of Apple’s third generation of foundation models, which Apple says it built in collaboration with Google. It has 20 billion parameters, but only 1 to 4 billion do work at any moment. Standard mixture-of-experts models pick fresh experts for every token, which forces all of them to sit in memory. Apple’s own research page says shuttling weights from flash storage to memory token by token is too slow, so Core Advanced picks its experts once per prompt and reselects them periodically while it writes.
Apple’s license bars reverse engineering, so Simon never opened the weights and watched the machine from outside. During an answer, the model pulls from one 4.95 GB file kept in flash, in chunks of 294,912 bytes each, about once every 32 tokens. Just 636 MiB of its pages remain resident on the Neural Engine, Apple’s dedicated AI chip. Simon marks the 32-token reselection, and the idea that those chunks are experts, as his own assumptions.
His comparison model is Gemma 4 26B A4B from Google, with 25.2 billion stored parameters and 3.8 billion active per token. Run at 4 bits, it held about 15 GiB shortly after a long request. Simon puts roughly half the difference down to bit width (Apple appears to use around 2 bits), and the rest on per-token routing, which makes a server keep every expert loaded.
Power draw follows the same pattern. Across 20 grade-school math problems, Apple’s model burned roughly one third as much energy as the best open build he tried. The Neural Engine pulled 4.2 watts on a long prompt, and the graphics processor stayed at its idle clock. On a laptop, where the model has to live beside a browser and an editor and run on battery, that frugality is the whole point.
Capability is what gets traded away. On 100 grade-school math problems, Apple’s model got 78 right, while three open builds got between 89 and 94. On 200 knowledge questions it landed in the middle. Apple’s developer guide already tells developers to avoid basic math, code, and logic puzzles, so the result confirms the design rather than exposing it.
Control is the second cost. An Apple engineer wrote that no API exists for pinning a model version, and the model changes with each operating system update. One developer’s test saw verdicts flip on 11 of 227 messages after an update, with identical code and prompt. An Apple staff member said developer adapters were removed in OS 27 for performance, so outsiders cannot tune it. On the test Mac the public API reports a window of 8,192 tokens, twice what Apple’s documentation still lists.
Simon also looks for Google’s fingerprints. Apple’s token counter agreed with Gemma 3’s on all 429 strings he tried, which points to a Google-derived tokenizer. His reading is that Google supplied training help and Apple built the flash layout, the chip programs, and the runtime around it. The technical report Apple promised for summer had not appeared by 25 October.
His closing argument is that Apple solved the memory problem well and should now release the weights with commercial use permitted, so developers can inspect and build on them. Until that happens, he recommends the model for short, fixed-shape work such as summaries, extraction, classification, and tags.
A team choosing between Apple’s model or an open alternative can settle it cheaply. Draft 100 questions drawn from your own workload, run both at zero temperature, and tally the questions only one answers correctly; Simon’s rule of thumb says a 20-question disagreement needs a 15-question lead to beat chance. Repeat the run after each OS update, because the model underneath shifts without notice.
Reported by AI Realist (airealist.ai), by Julien Simon, based on his measurements of 24 and 25 October 2026.
