What was done
yu314-coder reverse engineered the afmplus ifp container, weight codec and de-swizzle for the sparse mixture-of-experts model Apple ships with macOS. A controlled negative result documents where reconstruction still fails. No weights are distributed.
yu314-coder statically reverse engineered the on-device afmplus language model, the sparse mixture-of-experts backbone Apple ships with macOS. The repository publishes a teardown paper, decode and forward-pass code and a reproduction guide, and none of the Apple weights or tokenizer data. Twenty commits carry a Co-Authored-By: Claude trailer, with recovery, decode, validation and documentation all credited to the agents involved.
How it was done
The recovery pipeline copies the model asset that already exists on an Apple device you own, unpacks the container, reads its weight raster and program containers and rebuilds them as a single state dict, checked tensor by tensor against the shipped files. Every instrument used gets calibrated on a reference model before use, and several early positive claims were retracted in place once a statistic proved to carry no information, with the retractions kept as a record. A reproduce folder regenerates the decode from your own device without publishing Apple data.
Results
All 9.86 billion stored weights are recovered and validated, including the four-bit weight codec, the Apple Neural Engine de-swizzle and a 262,144 token byte-BPE tokenizer, and the expert router map is proven mathematically irrelevant to the ungated expert sum the runtime computes. A 44-layer forward pass runs but does not produce coherent text: a controlled negative result, published with the paper, localises the remaining gap to one shared convention invisible to every static statistic. Running the real model stays with the framework Apple ships.