OsaurusAI/Qwen-Image-2.1-Turbo-mflux-4bit

Research and evaluation use only — not for commercial use

These weights are a quantized derivative of Qwen/Qwen-Image-2.1-Turbo and are distributed under the Qwen Research License Agreement (LICENSE). The license permits use, reproduction and distribution for non-commercial purposes only, defined as research or evaluation. Any commercial use requires a separate license from the licensor (model-business@notice.qwencloud.com). By downloading or using these files you agree to the LICENSE.

Qwen-Image-2.1-Turbo — mflux 4-bit

An Osaurus 4-bit mflux quantization of Qwen/Qwen-Image-2.1-Turbo for Apple Silicon (9.9 GB, smallest; more composition drift between seeds). This is not an official Qwen release and is not endorsed by the Qwen team.

Turbo is the 8-step distilled checkpoint of Qwen-Image-2.1. One bundle does both text-to-image and reference editing (up to 10 ordered reference images); the vision encoder is included.

Variant	Size	Notes
8-bit	17 GB	closest to the original; recommended
6-bit	14 GB	balanced size and quality
4-bit	9.9 GB	smallest; more composition drift between seeds

Use in Osaurus

Requires Osaurus 0.25.21 or newer. Older versions treat this bundle as Qwen-Image-2.1 and sample it on that model's 40-step schedule, which gives degraded images.

Download it from the Image models tab, then generate or edit from chat or the image panel. Defaults: 8 steps on the model's fixed sampling grid, guidance 1.0 (no CFG), sizes in multiples of 32.

Sampling

The model samples on the fixed 8-node sigma grid stored in model_index.json (sample_sigmas), used unshifted with a terminal 0 and first-order Euler steps; there is no negative-prompt pass. A runtime that ignores this grid and applies Qwen-Image-2.1's shifted schedule produces degraded images. Stock mflux does not read the grid yet.

Changes from the original (LICENSE §3(b))

• Transformer and text/vision encoder weights re-saved with mflux and quantized to MLX affine 4-bit, group size 64 (modulation, timestep embedding and output norm kept in bf16).
• VAE re-saved in fp32, unquantized.
• model_index.json keeps the original 8-step sample_sigmas grid; no fine-tuning or other change to the model.

Generated images

The Qwen team has stated that "Outputs are not part of the licensed Materials. Users retain the rights to images and other content they generate using the model." This does not grant commercial use of the model itself; running the model remains limited to research and evaluation unless you hold a commercial license.

Attribution and license

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

"Qwen" is a trademark of its owner; it is used here only to describe the base model. No trademark license is granted or implied (LICENSE §5(b)).

Quantization: Osaurus · mflux (QwenImage21Edit save path) · 4-bit · eric@osaurus.ai