maxffarrell/amos-1-mini-CoreAI

AMOS-1 mini — Apple Core AI

An independently converted AMOS-1 mini decision model. Neural execution (encoder, trained projections, vision tower and projector) runs through Apple Core AI .aimodel assets. Upstream Python packing, pooling, probability normalization and decoding run on the host using PyTorch tensors. This is a decision model, not a text generator.

Use

Requires macOS 27 with the Core AI Python runtime. Tested with coreai-torch==0.4.3, PyTorch 2.14.0, NumPy, Pillow, tokenizers and huggingface_hub. The converter warns that this PyTorch version is newer than its validated range.

import asyncio
from coreai_mini import AmosMini

async def main():
    model = await AmosMini.load(".")
    print(await model.request({
        "state": "Please refund the duplicate charge.",
        "questions": {"refund": {
            "type": "noul", "instructions": "Is a refund requested?"
        }}
    }))

asyncio.run(main())

predict(state, questions, images=()) also accepts upstream internal questions (text, options) and PIL images, including image options. The request API accepts base64 images in the original AMOS format.

Capacity and verification

Batch one, 512 combined tokens, 224×224 images, FP32 weights. Requests exceeding capacity raise an error. This conversion does not provide the original 9,216-token capacity, state-cache acceleration, hosted API server, explain endpoint or benchmark claims. No iPhone or Neural Engine placement/performance has been verified. Core AI uses default specialization.

Nine reproducible smoke/parity cases cover noul, choice, score, multi, span, region, Unicode, 24 options, state images and image options. All tested choices, spans and region cells matched upstream. Maximum measured logit error was 0.000833. These smoke cases do not establish benchmark accuracy. validation.json contains exact results. runtime-smoke.json records inference with PyTorch Linear/Embedding/Conv2d execution disabled.

Reproduce and provenance

The original AMOS author is Zaid Mukaddam. AMOS source and weights are Apache-2.0; see LICENSE, SOURCE.json, and upstream AMOS-1 mini. The encoder derives from jhu-clsp/ettin-encoder-400m and the frozen vision tower from google/siglip2-base-patch16-224; dependency revisions are in DEPENDENCIES.json. Conversion adds no restrictions.

Run convert_mini.py --source PATH_TO_PINNED_AMOS_MINI --output OUTPUT, then validate_mini.py and smoke_mini.py from the original workspace layout. Original serving sources are included in serve/. SHA256SUMS covers release files.