D1A-E2B 路 MLX q8 (per-layer embeddings 4-bit)
The D1A-E2B v0.1 decision model (JohnP1/d1a-e2b) converted for Apple Silicon: the trained adapter merged into Gemma 4 E2B, linear layers and token embeddings at 8 bits, the large per-layer embedding table at 4 bits, and the pointer head in fp32 (head.safetensors). Typed questions in, a calibrated probability for every option out, one forward pass. Same System One API as Jev.
| This MLX build | PyTorch bf16 | |
|---|---|---|
| Memory after load / peak (60 questions) | 4.2 GB / 6.4 GB | ~10-12 GB / ~15 GB |
| Load time | 1.4 s | 16-20 s |
| 6-question request (M1 Max) | ~480 ms cold, ~400 ms warm | similar |
Parity against the fp32 reference, measured on 274 questions (201 decision-v7 development records, the playground presets, long states): mean |dp| 0.009, p95 0.030, max 0.084, 5 changed answers (all near-ties), accuracy 0.824 vs 0.811, ECE 0.062 vs 0.056. A pure 4-bit build failed the same gate (8.4% changed answers) and is not published.
Run it
pip install "d1a[serve] @ git+https://github.com/jonpol01/d1a@mlx-gemma4"
python -m d1a.serve --run JohnP1/d1a-e2b-mlx-q8 --port 8009
Apple Silicon only (MLX). Playground: https://github.com/jonpol01/d1a-playground
License
Apache-2.0. Base model: Gemma 4 by Google (Apache-2.0). Code: github.com/jonpol01/d1a, built on Kev by Jared Palmer (Apache-2.0).
- Downloads last month
- -
8-bit