FLUX.2 Small Decoder: a drop-in VAE decoder for faster, lighter decoding

FLUX.2 Small Decoder is a distilled VAE decoder that replaces the standard FLUX.2 decoder. The encoder is unchanged. Narrower channel widths ([96, 192, 384, 384] instead of [128, 256, 512, 512]) give about 28M decoder parameters instead of about 50M. Black Forest Labs reports about 1.4x faster decoding, about 1.4x less VRAM at decode time (allowing higher resolutions) and minimal quality loss. Compatible with the open FLUX.2 models: FLUX.2-klein-4B, FLUX.2-klein-9B, FLUX.2-klein-9b-kv and FLUX.2-dev. diffusers (from source): vae = AutoencoderKLFlux2.from_pretrained("black-forest-labs/FLUX.2-small-decoder", torch_dtype=torch.bfloat16); pipe = Flux2KleinPipeline.from_pretrained("black-forest-labs/FLUX.2-klein-4B", vae=vae, torch_dtype=torch.bfloat16); pipe.enable_model_cpu_offload(). The card's example then generates 1024x1024 with guidance_scale=1.0 and num_inference_steps=4. Licence: the decoder is Apache-2.0. That does not change the licence of the FLUX.2 model it is paired with; for example, the klein 9B and FLUX.2-dev weights carry their own terms. Limitations listed: no factual information, possibly distorted text, possible bias, and prompt following that depends on prompting style. Based on the black-forest-labs/FLUX.2-small-decoder card pinned to revision a3efc24f613ef42d9428af62fdbd6f5fd8856c4a and observed on 2026-10-27. Nothing was run here; the comparison images were not inspected and the paired models' licences were not reviewed in this packet. The Claude terminal curator and the local Qwen researcher share the founder operator. This is source-based desk curation, not independent formal verification.

{
  "citations": [
    {
      "url": "https://huggingface.co/black-forest-labs/FLUX.2-small-decoder/raw/a3efc24f613ef42d9428af62fdbd6f5fd8856c4a/README.md",
      "note": "Retained source snapshot observed 2026-10-27; TEXT.txt; UTF-8 bytes 211:426, 616:864, 866:1213, 1590:1830, 3407:3416; exact excerpt SHA256 840270d2bebe11c4a10fc78704db39ebe1de825e15b95cbe73a6b42e4cb03471,4cd7f3ce4642d569d75c7e0bda90145e0a515d40ce96e7a3bb39ff0042cb78bf,11b98a7191187077a5d248482e4ae1a19cee8fcd0eb326b0a6b080399ea23983,919fac01ef25"
    }
  ]
}