---
license: apache-2.0
license_link: LICENSE
base_model: vllm-sr/d3-mini
pipeline_tag: zero-shot-classification
library_name: coreai
tags:
  - vllm-sr
  - qwen3_5
  - decision
  - classification
  - system-one
  - decision-model
  - apple
  - coreai
  - ane
  - apple-neural-engine
  - fp16
  - tinkerbox
---

# d3-mini on Apple Neural Engine

Core AI conversion of [vllm-sr/d3-mini](https://huggingface.co/vllm-sr/d3-mini) for typed yes/no, choice, and score decisions over text and images, using [tinkerbox-ane](https://github.com/WeGoToMars/tinkerbox).

_This model has 4B parameters, for a samller 2B parameter model from the same family see [wegotomars/d3-nano](https://huggingface.co/wegotomars/d3-nano-CoreAI-ANE)._

d3-mini accepts prompts of up to 32,768 tokens. The Gated DeltaNet and full-attention layers and the vision tower run fully on the ANE. Tokenization, image preprocessing, position tables, token-embedding lookup, 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>=2.2.0`. Loading the model downloads only this Mac's bundle: its `checkpoint` folder carries the original model's code, tokenizer and configuration, and the token embeddings and vision position table the CPU uses.

```python
import tinkerbox_ane

with tinkerbox_ane.load("d3-mini") as model:
    result = model.system_one("I was charged twice. Please refund me today.", {
        "refund": {"type": "noul", "instructions": "Is this a refund request?"},
    })
    print(result)
```

Images can accompany text, or be used alone:

```python
from PIL import Image
import tinkerbox_ane

with tinkerbox_ane.load("d3-mini") 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)
```

d3's full-attention layers let every prompt token see the question, so questions cannot share a pass. A call with several questions runs one pass per question, in order, and warns with `SequentialQuestionsWarning`. 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 26, 2026 on Apple M5 Pro (`h17s`, 24 GiB), macOS 27.0.1, one question, warm.

| Metric | Result |
|---|---:|
| Top-answer agreement with FP32 text reference | 100% |
| Absolute probability drift, mean / maximum | 0.0008 / 0.025 |
| Official BF16 drift from FP32, mean / maximum | 0.0006 / 0.018 |
| Peak wired memory | 9.0 GiB |
| Peak Python memory | 0.76 GiB |
| 123-token text prompt | 86 ms |
| 198-token text prompt | 183 ms |
| 284-token text prompt | 276 ms |
| 358-token text prompt | 276 ms |
| 841-token text prompt | 0.65 s |
| 3,091-token text prompt | 2.47 s |
| 640 × 480 image prompt | 0.67 s |
| ANE energy, 123-token prompt | 0.52 J |
| ANE energy, 198-token prompt | 1.01 J |
| ANE energy, 284-token prompt | 1.56 J |
| ANE energy, 358-token prompt | 1.58 J |
| ANE energy, 841-token prompt | 3.68 J |
| ANE energy, 3,091-token prompt | 13.7 J |
| ANE energy, 640 × 480 image prompt | 2.88 J |

Prompts of 129 to 320 tokens take the chunked path, which keeps one set of weights resident to minimize memory instead of a wider, faster prompt function.

Original checkpoint: `vllm-sr/d3-mini@bd0090dd758d32ba59ef5c749910a8e87fc4391f`. Weights, original code, and converted assets retain the [Apache License 2.0](LICENSE). See [NOTICE](NOTICE).