anemll/jeff-ane-snake-coreai

Jeff Snake — Core AI build for the Apple Neural Engine

Unofficial Core AI conversion of a Snake-playing Jeff decision model (Jeff v1.3, a Qwen3.5-0.8B fine-tune) by Anemll. Each move is one decision, about 65 ms on an M5 Max. The packages run fully on the Apple Neural Engine (FP16, Core AI, p256_2k).

This is a demo, not a strong player: about 0.25 food per game in the 8-game self-play check (base: 0.12).

Source / code: github.com/Anemll/system1-ane (MIT): runtime, demo, training and conversion code.

What's here

config.json (package descriptor, not a transformers config), coreai/ (six chunk packages + head_readout, manifest.json), model/ (tokenizer, decision_config.json, readout, FP16 embeddings, transformers config.json), LICENSE, NOTICE. About 1.5 GB, exported not compiled.

Use

git clone https://github.com/Anemll/system1-ane.git && cd system1-ane
python scripts/download_model.py snake base
export COREAI_PYTHON=/path/to/coreai-sdk/bin/python
python forge.py compile --build ~/Models/jeff-ane/base/coreai
python forge.py compile --build ~/Models/jeff-ane/snake/coreai
python forge.py system1-ane-serve --model ~/Models/jeff-ane/base/model --build ~/Models/jeff-ane/base/coreai \
  --adapter snake=$HOME/Models/jeff-ane/snake/coreai --port 8796

Serve it as an adapter next to anemll/jeff-ane-base-coreai, which carries the full checkpoint system1-ane-serve --model needs.

Training

Rank-16 LoRA plus the readout head, trained locally on an Apple M5 Max with PyTorch MPS (2 epochs, about 5 minutes) using this project's own LoRA code, on Snake states generated and labelled by the repo's search oracle. No third-party dataset; no Unsloth or Colab weights are shipped.

Results (Apple M5 Max)

Held-out move accuracy (64 rows): base 0.281 → adapter 0.594 (MPS), 38/64 on the ANE build (matches PyTorch). Self-play (8 games): 0.25 food and 3.5 steps per game. All packages fully_ane. Details: docs/RESULTS.md.

Credits

• Jeff v1.3 by mstrasser: mstrasser/jeff-base (Apache-2.0).
• Qwen3.5-0.8B by the Qwen Team, Alibaba Cloud: Qwen/Qwen3.5-0.8B (Apache-2.0).
• Conversion to Core AI and the ANE runtime/demo: Anemll. Unofficial: not affiliated with or endorsed by Alibaba Cloud or the Jeff author.

License

Apache License 2.0 (LICENSE, attribution in NOTICE), the same license as Qwen3.5-0.8B and jeff-base v1.3, which these weights derive from. The demo and conversion code in Anemll/system1-ane is separate software under the MIT License.