Kwen-Image-2.1-Turbo-INT8

INT8 (W8A16) quantization of Qwen-Image 2.1 Turbo. Unofficial; not affiliated with Qwen or Alibaba. Built with Qwen.

Contents

Qwen-Image 2.1 Turbo (8 steps) for text-to-image, with the transformer weights in INT8 (W8A16). Every component keeps its original (diffusers / transformers) key names.

Path	Component	Format	Original
transformer/kwen-image-2.1-turbo-int8.safetensors	transformer	block linears <name>.weight int8 (out, in) + <name>.weight_scale fp16 (out,); every other tensor fp16	Qwen/Qwen-Image-2.1-Turbo/transformer/diffusion_pytorch_model-00001-of-00002.safetensors, Qwen/Qwen-Image-2.1-Turbo/transformer/diffusion_pytorch_model-00002-of-00002.safetensors
vae/diffusion_pytorch_model.safetensors	VAE	diffusers keys, fp16, time_conv removed	Qwen/Qwen-Image-2.1/vae/diffusion_pytorch_model.safetensors
text_encoder/	text encoder	Qwen3-VL-8B-Instruct-FP8 as published (fp8 e4m3, 128x128 block scales), without visual.* and lm_head; Qwen-Image's tokenizer and chat template	Qwen/Qwen3-VL-8B-Instruct-FP8

Processing

• Transformer: the bf16 shards merged into one file, keys unchanged. The 224 linears of the 32 blocks (attn.to_q/to_k/to_v/to_out.0, img_mlp.proj/gate_layer/out): 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. img_in, txt_in (with its norm), the timestep embedder, the shared modulation, norm_out, proj_out and the QK-norm scales are cast to fp16 (exact but for values below fp16's subnormal step). In the file each weight_scale directly precedes its weight.
• VAE: the F32 VAE of Qwen/Qwen-Image-2.1 (the Turbo repo's is the same, rounded to bf16), cast to fp16; the temporal time_conv layers, which never run on a single image, are removed.
• Text encoder: Qwen-Image's text encoder is Qwen3-VL-8B-Instruct; this is the official FP8 release of it, its tensors unchanged. The vision tower and lm_head are removed: Qwen-Image reads the last decoder layer's hidden states before the final norm. config.json is as published (it still describes the vision tower).

Licenses

Qwen-Image 2.1 Turbo (transformer, VAE, tokenizer): Qwen Research License, LICENSE; non-commercial (research and evaluation) use only. Attribution notice and the list of changes in NOTICE. Qwen3-VL-8B-Instruct-FP8 (text encoder): Apache-2.0, text_encoder/LICENSE.

Sources

• Qwen/Qwen-Image-2.1-Turbo at fc13ba1986fcb905854dc8de2f52eee395f49b5b (Qwen Research License)
  • transformer/config.json
  • processor/tokenizer.json
  • processor/tokenizer_config.json
  • processor/chat_template.jinja
  • processor/processor_config.json
  • LICENSE
• Qwen/Qwen-Image-2.1-Turbo at fc13ba1986fcb905854dc8de2f52eee395f49b5b (Qwen Research License)
  • transformer/diffusion_pytorch_model-00001-of-00002.safetensors
  • transformer/diffusion_pytorch_model-00002-of-00002.safetensors
• Qwen/Qwen-Image-2.1 at d26bb61231c349cf6b7896fa83353113880e1ba3 (Qwen Research License)
  • vae/diffusion_pytorch_model.safetensors
  • vae/config.json
• Qwen/Qwen3-VL-8B-Instruct-FP8 at 9cdc6310a8cb770ce18efaf4e9935334512aee45 (Apache-2.0)
  • config.json
  • generation_config.json
  • model.safetensors.index.json
  • model-00001-of-00002.safetensors
  • model-00002-of-00002.safetensors