README.md

---
license: agpl-3.0
library_name: mlx
tags:
  - object-detection
  - yolo
  - yolo26
  - mlx
  - apple-silicon
  - on-device
  - edge
pipeline_tag: object-detection
datasets:
  - coco
model-index:
  - name: yolo26x-mlx
    results:
      - task:
          type: object-detection
          name: Object Detection
        dataset:
          name: COCO val2017
          type: coco
        metrics:
          - type: mAP
            value: 0.567
            name: mAP@0.5:0.95
---

# YOLO26x (MLX)

Pure-MLX weights for **YOLO26x**, ready to run on Apple Silicon with
[`yolo-mlx`](https://github.com/thewebAI/yolo-mlx). No PyTorch at runtime,
no cloud calls, no waiting on someone else's API — everything stays on your Mac.

This is the largest, most accurate variant in the YOLO26 MLX family. Use it when accuracy is paramount and you can budget the extra compute per frame.

## Quickstart

```bash
pip install yolo-mlx huggingface_hub
```

```python
from huggingface_hub import hf_hub_download
from yolo26mlx import YOLO

weights = hf_hub_download("webAI-Official/yolo26x-mlx", "yolo26x.npz")
model = YOLO(weights)

results = model.predict("https://ultralytics.com/images/bus.jpg", conf=0.25)
results[0].save()
```

## Specs

| Variant | mAP@0.5:0.95 | FPS (M4 Pro) | Best for                     |
|---------|--------------|--------------|------------------------------|
| yolo26x | 56.7%        | 24          | Max accuracy, slower |

Other variants in this family:
[`yolo26n-mlx`](https://huggingface.co/webAI-Official/yolo26n-mlx) ·
[`yolo26s-mlx`](https://huggingface.co/webAI-Official/yolo26s-mlx) ·
[`yolo26m-mlx`](https://huggingface.co/webAI-Official/yolo26m-mlx) ·
[`yolo26l-mlx`](https://huggingface.co/webAI-Official/yolo26l-mlx)

## Requirements

- Apple Silicon Mac (M1, M2, M3, or M4)
- macOS 14.0+
- Python 3.10+

Intel Macs are not supported — the whole point of MLX is Apple Silicon native acceleration.

## What's in this repo

| File          | Description                                         |
|---------------|-----------------------------------------------------|
| `yolo26x.npz` | MLX-format weights, converted from the YOLO26x `.pt` checkpoint and verified shape-by-shape against the source. |
| `README.md`   | This card.                                          |

## Training data

Pretrained on [COCO](https://cocodataset.org/) (80 classes). For domain-specific
use cases, fine-tune on your own data — see the
[training guide](https://github.com/thewebAI/yolo-mlx/blob/main/GUIDE_TRAINING_BENCHMARK.md)
in the upstream repo.

## License

AGPL-3.0, inherited from upstream
[`thewebAI/yolo-mlx`](https://github.com/thewebAI/yolo-mlx).
Free to use, fork, modify, and ship for personal projects, research, and
prototypes. If you deploy this as a hosted service for real users, AGPL
requires you to publish your source under the same license.

## About webAI

[webAI](https://www.webai.com/) builds the sovereign AI platform — AI that runs
on your infrastructure, stays under your control, and compounds with your
knowledge. Every release here reflects a simple belief: **open models, owned
locally, coordinated intelligently, compound into something no centralized
system can match.**

🌐 [webai.com](https://www.webai.com/) · 💬 [community.webai.com](https://community.webai.com)