ryanhlewis/flux2-klein-4b-webgpu-lowbit

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. 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.

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:

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

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.