anemll/jeff-ane-tetris-coreai

Jeff Tetris — Core AI build for the Apple Neural Engine

Unofficial Core AI conversion of a Tetris-playing Jeff decision model (Jeff v1.3, a Qwen3.5-0.8B fine-tune) by Anemll. Each Tetris move is one forward pass: Jeff scores every legal placement and the demo plays the most probable one. The packages run fully on the Apple Neural Engine (FP16, Core AI).

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

What's here

Path	What
config.json	Package descriptor (not a transformers config)
coreai/	chunk_L00-03 … chunk_L20-23 + head_readout .aimodel packages and manifest.json (p256_2k, FP16, context 2048)
model/	Tokenizer, chat template, decision_config.json, 255-way readout.safetensors, FP16 embedding table, transformers config.json
LICENSE, NOTICE	Apache-2.0 and attribution

About 1.5 GB. The packages are exported, not compiled: compile once per Mac. system1-ane-serve --model needs the full base checkpoint, so serve this build as an adapter next to anemll/jeff-ane-base-coreai.

Use

git clone https://github.com/Anemll/system1-ane.git && cd system1-ane
python scripts/download_model.py tetris base          # -> ~/Models/jeff-ane/{tetris,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/tetris/coreai
python forge.py system1-ane-serve --model ~/Models/jeff-ane/base/model --build ~/Models/jeff-ane/base/coreai \
  --adapter tetris=$HOME/Models/jeff-ane/tetris/coreai --port 8796
# open http://127.0.0.1:8796/ and pick the Tetris tab

Requires Apple Silicon and macOS with Core AI. Measured on an Apple M5 Max, macOS 27.2 (bonded ANE compile mode 1).

Training

Rank-16 LoRA (alpha 32) plus the readout head, trained locally on an Apple M5 Max with PyTorch MPS using this project's own (hand-rolled) LoRA code — no Unsloth, no Colab. The shipped weights are round 02 of a bounded overnight continuation (2026-10-08) on 2,048-row mixtures of teacher games, noisy-exploration games, synthetic-garbage recovery states and teacher corrections collected from Jeff's own boards. All labels come from the El-Tetris heuristic by Islam El-Ashi (2011), an improvement on Pierre Dellacherie's algorithm (article); the games are generated by the repo. No third-party dataset. Round 02 was chosen on 8 development seeds.

Results (Apple M5 Max, greedy, 150-piece cap, 32 benchmark seeds)

Policy	Mean pieces	Mean lines	Reached cap
Previous Tetris adapter	59.84	12.41	0/32
This model (round 02)	149.75	53.75	30/32
El-Tetris teacher	150	57.69	32/32

• The 32 seeds had been evaluated before (a repeated benchmark, not a pristine holdout). On 8 fresh seeds it reached the cap 8/8 with 55.1 lines on average.
• Warm latency about 293 ms/move on average (p50 252 ms, p95 379 ms; ~1,400-token prompts, 6 × 256-row ANE calls).
• Held-out move accuracy vs the teacher (256 rows, MPS): 0.684 → 0.746. ANE vs PyTorch parity: 4/4 argmax, mean KL 2.0e-3. All packages report fully_ane.
• Browser demo games vary (no piece cap): in 4 recorded games two topped out (at 66 and 163 pieces) and two were still alive at 224 and 219 pieces after 110 s.

Limitations

• A narrow game policy, not a general assistant. Prompts must use the demo's exact board text and option prose.
• FP16 ANE packages: near-tie options can flip versus FP32.
• Verified only on M5 Max / macOS 27.2. Other Macs must compile locally (M6+ uses compile mode 2).

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).
• El-Tetris by Islam El-Ashi (2011), an improvement on Pierre Dellacherie's algorithm: https://imake.ninja/el-tetris-an-improvement-on-pierre-dellacheries-algorithm/ (Tetris labels and reference player).
• 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.