| --- |
| license: apache-2.0 |
| base_model: black-forest-labs/FLUX.2-klein-4B |
| base_model_relation: quantized |
| pipeline_tag: text-to-image |
| library_name: kaminos |
| tags: |
| - webgpu |
| - text-to-image |
| - flux |
| - quantized |
| - browser |
| --- |
| |
| # FLUX.2 [klein] 4B for WebGPU |
|
|
| Quantized weights of [FLUX.2 [klein] 4B](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B), packed for text-to-image generation entirely in the browser with the [Kaminos](https://github.com/lyonsno/kaminos) WebGPU inference kit. The page downloads about 3.2 GB once (gzip copies of everything except the embedding table, from which it fetches only the rows a prompt uses), caches it, and generates on the visitor's own GPU. No server-side compute is involved. |
|
|
| **[Try it in your browser](https://lyonsno.github.io/kaminos/inference-kit/klein/)** · [Source](https://github.com/lyonsno/kaminos/tree/main/models/flux2-klein) |
|
|
| In Chrome, a 512 × 512 image (4 steps) takes 6–7 seconds on an Apple M4 Max and about 14 seconds on a 16 GB Apple M2 Pro. The page also offers 768 and 1024, and frees each stage's working memory before the next, so a 16 GB Mac generates at 1024 × 1024 without swapping. |
|
|
| ## Contents |
|
|
| | Folder | Component | Format | Size | |
| | --- | --- | --- | --- | |
| | `te/` | Qwen3-4B text encoder, layers 0–26 (the pipeline reads hidden states 9, 18 and 27) | int4 weights, group 128, affine; norms f16 | 1.45 GB | |
| | `te/te-embed.bin` | Token-embedding table, read one row per token with HTTP range requests | f16 | 0.78 GB | |
| | `te/tokenizer.json` | Qwen2 byte-level BPE tokenizer | JSON | 11 MB | |
| | `dit/` | Rectified-flow transformer | block linears int4 (group 128, affine); embedders, modulation and output projections int4 (group 64, affine); norms f16 | 2.07 GB | |
| | `vae/` | VAE decoder and latent batch-norm statistics | f16; statistics f32 | 0.1 GB | |
| | `dit-q3/` | Transformer, "Weathered" level | block linears 3-bit (group 64, affine; per-group range chosen to minimize squared error); the rest as in `dit/` | 1.64 GB (gzip only) | |
| | `dit-q2/` | Transformer, "Off the Rails" level | block linears 2-bit (group 64, as above); the rest as in `dit/` | 1.16 GB (gzip only) | |
|
|
| The page's jank-level selector swaps between `dit/` (Mint), `dit-q3/` (Weathered: looks about the same) and `dit-q2/` (Off the Rails: blurry and melted, still recognizable); the text encoder, decoder and tokenizer are shared. 3-bit codes are packed 32 per three 32-bit words, 2-bit codes 16 per word. |
|
|
| Each folder has a `manifest.json` listing every tensor's shape, format, byte offset, and the SHA-256 of each file. Every bundle and the tokenizer also have a gzip copy (`.gz`, listed in the manifest), which is what the page downloads; the embedding table stays uncompressed for range requests. Quantization is weight-only. Activations and the residual stream stay in floating point. |
|
|
| ## Changes from the original |
|
|
| - Weights quantized and repacked into per-block files as described above. |
| - Text encoder truncated to the layers the FLUX.2 [klein] pipeline uses. |
| - Projections that share an input are fused (for example query, key and value). |
| - 3×3 VAE convolution weights reordered for channels-last execution. |
|
|
| ## License |
|
|
| Apache License 2.0, the same license as the original model. FLUX.2 [klein] 4B is by Black Forest Labs. Its text encoder is Qwen3-4B by the Qwen team, also Apache 2.0. |
|
|
