---
title: Qwen3-8B-DFly-MLX
canonical_url: "https://www.modelscope.cn/models/mlx-community/Qwen3-8B-DFly-MLX"
md_url: "https://www.modelscope.cn/models/mlx-community/Qwen3-8B-DFly-MLX.md"
repository: mlx-community/Qwen3-8B-DFly-MLX
last_updated: 2026-08-01
license: apache-2.0
model_type:
  - qwen3
architectures:
  - DFlyDraftModel
base_model:
  - Qwen/Qwen3-8B
base_model_relation: finetune
parameters: 1.1B
library_name:
  - mlx
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 14
stars: 0
tags:
  - mlx
  - qwen3
  - speculative-decoding
  - dfly
  - apple-silicon
---

# Qwen3-8B-DFly-MLX

> Qwen3-8B-DFly-MLX - mlx-community 在 ModelScope 开源的模型。MLX conversion of AngelSlim/Qwen3-8B-DFly-Block8 for native DFly speculative decoding on Apple Silicon.

mlx-community/Qwen3-8B-DFly-MLX 是 ModelScope 魔搭社区上的 1.1B 参数机器学习模型，采用 apache-2.0 许可，基于 Qwen/Qwen3-8B 构建。

- **Repository**: mlx-community/Qwen3-8B-DFly-MLX
- **License**: apache-2.0
- **Parameters**: 1.1B
- **Base model**: Qwen/Qwen3-8B
- **Tags**: mlx, qwen3, speculative-decoding, dfly, apple-silicon
- **Downloads**: 14
- **Stars**: 0
- **Last updated**: 2026-08-01

Source: https://www.modelscope.cn/models/mlx-community/Qwen3-8B-DFly-MLX

---

# Qwen3-8B DFly MLX

MLX conversion of [`AngelSlim/Qwen3-8B-DFly-Block8`](https://huggingface.co/AngelSlim/Qwen3-8B-DFly-Block8) for native DFly speculative decoding on Apple Silicon.

The conversion uses upstream revision `5712926`, the last revision with the correct Qwen3-8B metadata. The current upstream `config.json` describes an incompatible 80-layer target, so `dfly-mlx` pins and validates the known-good revision.

## Use

```sh
pip install dfly-mlx
hf download mlx-community/Qwen3-8B-DFly-MLX \
  --local-dir ./models/qwen3-8b-dfly-mlx

dfly-mlx generate \
  --draft ./models/qwen3-8b-dfly-mlx \
  --prompt "Write a quicksort in Python." \
  --max-tokens 256
```

The target defaults to `mlx-community/Qwen3-8B-4bit`. Generation is greedy and uses Qwen3 no-thinking mode.

## Verification status

Checkpoint conversion, loading, and generation were tested on Apple Silicon with MLX 0.32.0. Every emitted draft proposal is accepted by the target verifier. As with the underlying DFlash runtime, output can differ from pure autoregressive generation near tied logits because MLX dispatch differs between single-token and block verification.

Source, reproducible benchmark command, and attribution: [will702/dfly-mlx](https://github.com/will702/dfly-mlx).

## License

Apache-2.0. The original AngelSlim model license is included as `License_AngelSlim_model_and_dataset.txt`.