---
title: libwaifu-qwen-image-2.1-turbo
canonical_url: "https://www.modelscope.cn/models/ling0322/libwaifu-qwen-image-2.1-turbo"
md_url: "https://www.modelscope.cn/models/ling0322/libwaifu-qwen-image-2.1-turbo.md"
repository: ling0322/libwaifu-qwen-image-2.1-turbo
last_updated: 2026-10-14
license: other
pipeline_tag: text-to-image-synthesis
tasks:
  - text-to-image-synthesis
base_model:
  - Qwen/Qwen-Image-2.1-Turbo
base_model_relation: finetune
parameters: 3.3B
tensor_type:
  - F8_E4M3
  - F16
  - F32
library_name:
  - safetensors
language:
  - en
downloads: 4
stars: 0
tags:
  - libwaifu
  - qwen-image
---

# libwaifu-qwen-image-2.1-turbo

> libwaifu-qwen-image-2.1-turbo - ling0322 在 ModelScope 开源的模型。Qwen-Image 2.1 Turbo (fp8) — the libwaifu package

ling0322/libwaifu-qwen-image-2.1-turbo 是 ModelScope 魔搭社区上的 3.3B 参数text-to-image-synthesis模型，采用 other 许可，基于 Qwen/Qwen-Image-2.1-Turbo 构建。

- **Repository**: ling0322/libwaifu-qwen-image-2.1-turbo
- **License**: other
- **Tasks**: text-to-image-synthesis
- **Parameters**: 3.3B
- **Base model**: Qwen/Qwen-Image-2.1-Turbo
- **Tags**: libwaifu, qwen-image
- **Downloads**: 4
- **Stars**: 0
- **Last updated**: 2026-10-14

Source: https://www.modelscope.cn/models/ling0322/libwaifu-qwen-image-2.1-turbo

---

# Qwen-Image 2.1 Turbo (fp8) — the libwaifu package

**Qwen-Image-2.1-Turbo**, converted to the package format [libwaifu](https://github.com/ling0322/libwaifu)
reads: the 8-step accelerated checkpoint of Qwen-Image 2.1, with the matrices stored as fp8.

> **By downloading these files you agree to the [Qwen RESEARCH LICENSE
> AGREEMENT](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo/blob/main/LICENSE)**, a copy of which
> is distributed here as `LICENSE`. **The agreement is non-commercial (research and evaluation) use
> only** — see `LICENSE` for the exact terms, including how to request a commercial license.
>
> **This is a modified copy** — see `NOTICE` for what was changed — and it is **not an official
> Qwen product and is not endorsed by Alibaba or the Qwen team**. This project is Built with Qwen.
> The original weights are at
> [Qwen/Qwen-Image-2.1-Turbo](https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo).

## What is here

| package | on disk | what it is |
|---|---|---|
| `qwen-image-2.1-turbo-fp8.yaml` and its parts | 17.1 GB in ten parts, none over 2 GB | the matrices as E4M3 with one scale for the whole tensor (`fp8_tensor_scale`), everything else float16 |

It carries the whole model: the 7B single-stream DiT, the language half of a Qwen3-VL-8B text
encoder (read at its last layer), and the Wan-2.2-style autoencoder (encoder and decoder) — 64
latent channels, an alpha channel out. The Qwen3-VL **vision tower** is included too (float16,
1.15 GB, under `qwen_image.vision.*`) for image editing; the runtime does not read it yet.
Nothing else has to be fetched to draw. fp8 needs an sm_80 or newer card.

## Drawing with it

```bash
waifu webui -m qwen-image:2.1-turbo-fp8
```

The manifest fixes the checkpoint's **eight sampling sigmas** (`sampler_sigmas`), so the number of
steps you ask for is ignored: it is always eight, at no guidance. Qwen-Image 2.1's own packages
bend a schedule by picture size instead; this one does not.

## How close it is to the original

Against diffusers' `QwenImage21Pipeline` on the same weights, 256x256, eight steps, with every
input taken off one real trajectory:

| | fp8 package |
|---|---|
| encoder, last layer before the norm (against float32) | 1.12e-1 |
| velocity at step five | 2.4e-2 |
| decoder, RGBA (against float32) | 8.0e-4 — it is not quantized |
| the whole eight-step walk | 1.19e-1 |

The text encoder's last layer is the sensitive part, as with Qwen-Image 2.1's own fp8 package.

## Attribution

Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi
Laboratory Technology Co., Ltd. All Rights Reserved. This project is Built with Qwen.