System 1 ANE — Core AI builds for the Apple Neural Engine
Small decision models (one forward pass, one decision, no generated text) converted to Core AI .aimodel packages that run on the Apple Neural Engine. Each model lives in its own folder with its own config.json and model card.
| Model | Folder | Base | Task | Notes |
|---|---|---|---|---|
| Snake, stock Qwen3.5-0.8B + Unsloth-trained LoRA and decision head | snake-stock-qwen3.5-unsloth/ | unsloth/Qwen3.5-0.8B | Snake move decision | 304-token prefill window, 7/7 packages fully_ane, ~91 ms per decision on an M5 Max |
More models may be added as further folders. This repository is a collection of Core AI builds, not a transformers checkpoint: transformers AutoModel cannot load it. The root config.json is a repository descriptor (it lists the models); each model folder has its own package descriptor.
Download one model
from huggingface_hub import snapshot_download
path = snapshot_download(
repo_id="anemll/system1-base-unsloth",
allow_patterns=["config.json", "snake-stock-qwen3.5-unsloth/*"],
local_dir="~/Models/system1-base-unsloth",
)
Source / code: github.com/Anemll/system1-ane (MIT): runtime, demo, training and conversion code.
See the model card in each folder for how to compile and run it.
License
Apache License 2.0 (LICENSE, attribution in NOTICE), the same license as Qwen3.5-0.8B and Unsloth, which these weights derive from. Unofficial: not affiliated with or endorsed by Alibaba Cloud or Unsloth.