d3-mini — Apple Core AI conversion

Community conversion of vllm-sr/d3-mini (part of the Decision 3.0 collection) to an Apple Core AI .aimodel bundle, runnable on macOS 27 on Apple Silicon.

Contains the text decision backbone, the original 255-way answer readout, and the vision encoder. Choice, yes/no (noul) and score questions run through Core AI; tokenization, media decoding and result formatting run on the host. No text is generated.

Use

hf download ahmedandaloes/d3-mini-CoreAI --local-dir d3-mini-CoreAI
cd d3-mini-CoreAI
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

import asyncio
from coreai_d3 import D3CoreAI

async def main():
    model = await D3CoreAI.load(".")
    print(await model.system_one(
        state="The customer was charged twice.",
        questions={"route": {"type": "choice",
                   "instructions": "Which team should handle this?",
                   "criteria": {"billing": "Payments", "shipping": "Delivery"}}},
    ))

asyncio.run(main())

Capacity and changes from upstream

One question per forward pass; system_one processes several questions sequentially.
512 combined tokens (visual + prompt text). Oversized requests raise an error — nothing is truncated.
Images and sampled video frames are resized to 224×224, including aspect-ratio changes.
The original BF16 backbone weights and FP32 decision head are retained.
Linear attention stays causal; full attention remains noncausal, as upstream specifies.

Numerical drift

Measured on 6 local smoke cases against the upstream BF16 reference: all selected answers matched — True; largest absolute probability difference 0.001387. Small smoke tests, not a quality benchmark.

Files

decision.aimodel — text backbone + readout head
vision.aimodel — vision encoder
coreai_d3.py, d3_format.py, d3_runtime.py — host adapter and upstream prompt/answer contract
manifest.json — source repo, pinned revision, sequence length, precision, exported I/O names

Independent community conversion — not an official Apple or vLLM release. Converted on macOS 27.0.1 with coreai-core 1.0.0b3 / coreai-torch 0.4.3.