pushkarmoi/flux2-klein-9b-int8

FLUX.2 [klein] 9B INT8

Contents

Unofficial; not affiliated with Black Forest Labs.

FLUX.2 [klein] 9B (distilled, 4 steps) for text-to-image, with the transformer weights in INT8 (W8A16). Every component is in BFL's original key layout.

Path	Component	Format	Original
transformer/flux-2-klein-9b-int8.safetensors	transformer	every linear <name>.weight int8 (out, in) + <name>.weight_scale fp16 (out,); QK-norm scales bf16	black-forest-labs/FLUX.2-klein-9b-fp8/flux-2-klein-9b-fp8.safetensors
vae/ae.safetensors	VAE (decoder + bn only)	BFL AutoEncoder keys, bf16	black-forest-labs/FLUX.2-klein-9B/vae/diffusion_pytorch_model.safetensors
text_encoder/	text encoder	Qwen3-8B-FP8 as published (fp8 e4m3, 128x128 block scales) + tokenizer	Qwen/Qwen3-8B-FP8

Processing

• Transformer: quantized from BFL's FP8 checkpoint (each fp8 e4m3 linear times its per-tensor weight_scale; its bf16 layers as they are), keys unchanged. Symmetric int8 in [-127, 127], one fp16 scale per output channel (scale = absmax / 127, rounded to fp16 before the weights are quantized), round to nearest, no calibration (torchao choose_qparams_affine / quantize_affine). Dequantize as W = weight * weight_scale[:, None]; activations stay fp16.
• VAE: the repo's diffusers VAE (AutoencoderKLFlux2) with its keys mapped back to BFL's AutoEncoder layout (the inverse of diffusers' scripts/convert_flux2_to_diffusers.py); decoder and bn only, values unchanged. BFL's own ae.safetensors is only in the gated FLUX.2 [dev] repo.
• Text encoder: copied unchanged. FLUX.2 [klein] reads the hidden states of Qwen3 layers 9, 18 and 27.

Licenses

FLUX.2 [klein] 9B (transformer, from its FP8 release; VAE): FLUX Non-Commercial License (the same in both repos), LICENSE.md; non-commercial use only. Attribution Notice in NOTICE. Qwen3-8B-FP8 (text encoder): Apache-2.0, text_encoder/LICENSE.

Sources

• black-forest-labs/FLUX.2-klein-9B at 92196c8e11f7b6cf2b7493e037d8c5345c559216 (FLUX Non-Commercial License)
  • vae/diffusion_pytorch_model.safetensors
  • vae/config.json
  • LICENSE.md
• Qwen/Qwen3-8B-FP8 at 220b46e3b2180893580a4454f21f22d3ebb187d3 (Apache-2.0)
  • config.json
  • generation_config.json
  • model.safetensors.index.json
  • model-00001-of-00002.safetensors
  • model-00002-of-00002.safetensors
  • tokenizer.json
  • tokenizer_config.json
  • vocab.json
  • merges.txt
  • LICENSE
• black-forest-labs/FLUX.2-klein-9b-fp8 at 902d9d510b51533e07729f19211414a3648b77d2 (FLUX Non-Commercial License)
  • flux-2-klein-9b-fp8.safetensors