anemll/jeff-ane-base-coreai

Jeff v1.3 base — Core AI build for the Apple Neural Engine

Unofficial Core AI conversion of mstrasser/jeff-base (Jeff v1.3, a Qwen3.5-0.8B fine-tune) by Anemll. No retraining: the weights are Jeff's, converted to six 4-layer .aimodel chunks plus the 255-way readout head, all fully_ane on an M5 Max. You send a situation and named options; one forward pass returns a calibrated probability per option (about 65 ms for a short decision).

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/	Core AI packages and manifest.json (p256_2k, FP16, context 2048)
model/	Tokenizer, decision_config.json, readout, FP16 embeddings, transformers config.json and model.safetensors (the unmodified jeff-base v1.3 checkpoint that system1-ane-serve --model loads)
LICENSE, NOTICE	Apache-2.0 and attribution

About 3.2 GB, exported not compiled. This is the server's base build; the Snake and Tetris builds load beside it with --adapter (snake, tetris).

Use

git clone https://github.com/Anemll/system1-ane.git && cd system1-ane
python scripts/download_model.py base
export COREAI_PYTHON=/path/to/coreai-sdk/bin/python
python forge.py compile --build ~/Models/jeff-ane/base/coreai
python forge.py system1-ane-serve --model ~/Models/jeff-ane/base/model --build ~/Models/jeff-ane/base/coreai --port 8796
# open http://127.0.0.1:8796/ (routing panel)

Jeff v1.3 is meant to be used with task adapters; on its own it is weak on unfamiliar tasks with long option lists (see the jeff-base card).

Results (Apple M5 Max, macOS 27.2)

Prefill parity vs PyTorch FP32 on real Jeff prompts: KL ≤ 8.3e-4, same argmax except one near-tie at 2,018 tokens. About 64 ms per short decision (one 256-row call). 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.