---
license: other
language:
- en
pipeline_tag: text-to-image
tags:
- flux
- flux2
- webgpu
- onnx
- onnxruntime
- image-generation
- text-to-image
- lowbit
base_model: black-forest-labs/FLUX.2-klein-4B
library_name: onnxruntime
---

# FLUX.2 Klein 4B WebGPU Low-Bit Bundle

This is the browser model bundle for the static FLUX.2 WebGPU app at
[ryanhlewis/flux2-webgpu](https://huggingface.co/spaces/ryanhlewis/flux2-webgpu).
It is derived from `black-forest-labs/FLUX.2-klein-4B` and the ONNX WebGPU q4
bundle by `MarkShark2/flux2-klein-4b-onnx-webgpu-q4`, with an additional custom
low-bit WebGPU transformer runtime payload used by
[ryanhlewis/flux2-webgpu](https://github.com/ryanhlewis/flux2-webgpu).

## Contents

| Component | Size |
| --- | ---: |
| Custom low-bit transformer assets | 7.27 GB / 6.77 GiB |
| Upstream q4 text encoder and ONNX transformer fallback | 5.13 GB / 4.77 GiB |
| VAE decoder/encoder, tokenizer, config overlay | 0.18 GB / 0.17 GiB |
| Precomputed default prompt context/projection | 0.01 GB / 0.01 GiB |
| Total staged model repository payload | 12.58 GB / 11.72 GiB |

The default static UI uses the custom low-bit transformer path. Arbitrary prompts
require the q4 text encoder. The q4 ONNX transformer files are included as a
fallback path, but the app defaults to the custom low-bit transformer runtime.

## Static Browser Use

Serve the app files from the GitHub repository or the linked Static Space and set:

```js
globalThis.FLUX2_MODEL_BASE_URL = "https://huggingface.co/ryanhlewis/flux2-klein-4b-webgpu-lowbit/resolve/main";
globalThis.FLUX2_RUNTIME_BASE_URL = globalThis.FLUX2_MODEL_BASE_URL;
globalThis.FLUX2_CUSTOM_KERNEL_BASE_URL = `${globalThis.FLUX2_MODEL_BASE_URL}/custom_lowbit`;
```

No Python, Gradio server, CUDA, or ROCm runtime is required for generation in the
Static Space. The user's browser downloads model files and runs inference locally
through WebGPU.

## Benchmarks

Local headed Chrome/Edge WebGPU benchmarks on the development machine, after
background preparation, with result cache disabled, default prompt, fixed seeds,
custom low-bit WebGPU backend, and exact full-resolution render/decode:

| Size | Browser elapsed | Transformer | VAE | Notes |
| ---: | ---: | ---: | ---: | --- |
| 256x256 | 0.817s | 0.603s | 0.213s | exact |
| 512x512 | 4.438s | 2.922s | 1.514s | exact |
| 768x768 | 10.723s | 8.009s | 2.712s | exact |
| 1024x1024 | 45.434s | 24.108s | 21.324s | exact, VAE-heavy |

These are browser WebGPU numbers, not PyTorch/CUDA numbers. First load will also
include model download and browser cache population time.

## Example Outputs

Default robot prompt, seed `123`, quality/custom WebGPU path:

| 256 | 512 | 1024 |
| --- | --- | --- |
| ![256 example](examples/example-256.png) | ![512 example](examples/example-512.png) | ![1024 example](examples/example-1024.png) |

## Memory Expectations

WebGPU does not expose exact VRAM usage to JavaScript, so the app reports
observable browser memory and known tensor/buffer sizes. The warmed custom
transformer path accounts for roughly 3.7 GB of loaded GPU/buffer data. Exact
256 and 512 are realistic on an 8 GB WebGPU adapter when browser memory limits
allow it. Exact 768 and 1024 are memory-heavy; the 1024 benchmark completed with
browser heap around 4.2-4.3 GiB during VAE decode.

## License

This bundle is derived from upstream FLUX.2 Klein 4B assets. Use and
redistribution must comply with the upstream model license and applicable
Hugging Face terms.