Qwen-Image-2.1-Turbo (unmodified mirror, DiT + pipeline config)

An unmodified, byte-identical mirror of the parts of Qwen/Qwen-Image-2.1-Turbo that differ from the base model, at upstream revision d65dbc9a7e8f6b5479e33dee6030eaab2a906509 (2026-10-09). Every file was checked against the upstream LFS SHA-256 / git blob hash before upload.

It is the weight source for the Turbo tier of qwen-image21-swift, a Swift/MLX port of the model for Apple Silicon. The port reads this diffusers layout directly; no conversion is applied.

Licence: research / evaluation only

Qwen-Image-2.1-Turbo is licensed under the Qwen RESEARCH LICENSE AGREEMENT (included as LICENSE). Use is limited to non-commercial research or evaluation; commercial use requires a separate licence from Qwen (model-business@notice.qwencloud.com). This mirror is redistributed under §3 of that agreement, with the agreement and the attribution notice (NOTICE) included. Nothing here grants any right beyond that agreement.

Qwen is licensed under the Qwen RESEARCH LICENSE AGREEMENT, Copyright (c) 2026 Hangzhou Tongyi Laboratory Technology Co., Ltd. All Rights Reserved.

Contents

Path	What
transformer/	the Turbo DiT — the 2.1 7B single-stream block-causal transformer re-weighted for 8 steps (bf16 safetensors, 2 shards)
model_index.json	pipeline index carrying sample_sigmas, the 8-step sampling grid the checkpoint was distilled on (diffusers PR #14950)
scheduler/	FlowMatchEulerDiscreteScheduler config with dynamic shifting and the terminal stretch switched off — the grid is used verbatim
LICENSE, NOTICE	the agreement, attribution

Not included — load them from where they already live:

• vae/: upstream's Turbo vae/ is the base model's VAE cast to bf16 (same 238 tensors; sampled weights differ from the base fp32 by bf16 rounding only). Use the fp32 one from xocialize/Qwen-Image-2.1 (vae/) or upstream Qwen/Qwen-Image-2.1.
• text_encoder/ and processor/: byte-identical to the Apache-2.0 Qwen/Qwen3-VL-8B-Instruct (750/750 tensors, same chat template).

model_index.json still names every component, so diffusers users pass the missing ones explicitly, e.g. QwenImage21Pipeline.from_pretrained("xocialize/Qwen-Image-2.1-Turbo", vae=AutoencoderKLQwenImage21.from_pretrained("xocialize/Qwen-Image-2.1", subfolder="vae"), text_encoder=Qwen3VLForConditionalGeneration.from_pretrained("Qwen/Qwen3-VL-8B-Instruct", torch_dtype=torch.bfloat16), processor=Qwen3VLProcessor.from_pretrained("Qwen/Qwen3-VL-8B-Instruct"), torch_dtype=torch.bfloat16). The saved 8-step schedule loads automatically with diffusers ≥ PR #14950; num_inference_steps does not override it.

For the model itself, see the upstream model card.