wegotomars/d1-3B-CoreAI-ANE

d1-3B on Apple Neural Engine

Core AI conversion of LiquidAI/d1-3B for typed yes/no, choice, and score decisions over text and images, using tinkerbox-ane.

The language backbone and vision tower run fully on the ANE. Tokenization, image preprocessing, token-embedding lookup, the FP32 multimodal projector, and answer readout run on the CPU while your GPU stays idle and cool 😎❄️.

Requires macOS 27, Apple Silicon M4 or newer, and Python 3.12 or 3.13.

Usage

Install tinkerbox-ane>=2.1.0. Loading the model downloads only this Mac's bundle: its checkpoint folder carries the upstream code, tokenizer and configuration, and the token embeddings, multimodal projector and vision position table the CPU uses.

Evaluate independent questions together, sharing computation over the input:

from tinkerbox_ane import d1

with d1.load() as model:
    result = model.system_one("I was charged twice. Please refund me today.", {
        "refund": {"type": "noul", "instructions": "Is this a refund request?"},
        "urgent": {"type": "noul", "instructions": "Does this need an urgent response?"},
    })
    print(result)

Images can accompany text, or be used alone:

from PIL import Image
from tinkerbox_ane import d1

with d1.load() as model, Image.open("receipt.jpg") as image:
    result = model.system_one(None, {
        "receipt": {"type": "noul", "instructions": "Is this a receipt?"},
    }, images=[image])
    print(result)

Each result's usage reports its tokens, Neural Engine time, and counted FLOPs.

Supported architectures

Architecure	SoC name
h16g	Apple M4
h16s	Apple M4 Pro
h16c	Apple M4 Max
h17g	Apple M5
h17s	Apple M5 Pro
h17c	Apple M5 Max
h17d	Apple M5 Ultra
h18g	Apple M6

Benchmarks

Measured October 19, 2026 on Apple M5 Pro (h17s, 24 GiB), macOS 27.0.1.

Metric	Result
Top-answer agreement with FP32 text reference	100% (204/204)
Absolute probability drift, mean / maximum	0.000665 / 0.009730
One text question, 26–251 tokens: median / p95	39.3 / 79.9 ms
4,096-token text prompt: median	1.49 s
32,768-token text prompt	51.8 s
Two parallel questions over one short text: median / p95	39.2 / 39.7 ms
One image prompt: median / p95	216.8 / 220.1 ms

Upstream checkpoint: LiquidAI/d1-3B@051bcc464b01b9f92942b364d9586b0ef5912432. Weights, upstream code, and converted assets retain the LFM Open License v1.0. See NOTICE.