- Unsloth published GGUF and FP8 quantizations of Qwen-Image-2.1 on Hugging Face on September 21, making the 7B image model runnable on consumer hardware.
- Unsloth's documentation states memory figures are estimates, not tested minimums, and recommends FP8 over GGUF for GPU-equipped users.
- The base Qwen-Image-2.1 model is governed by a Qwen Research License Agreement dated September 20 that permits only non-commercial research and evaluation use.
- Derivative works must display 'Built with Qwen' attribution and cannot use 'Qwen' as a primary product name under the license terms.
Quantizations target consumer hardware
Unsloth, a third-party optimization developer, uploaded GGUF and FP8 quantized versions of Alibaba's Qwen-Image-2.1 to Hugging Face on September 21, 2026. The GGUF repository was created at 13:04 UTC that day, following an FP8 upload earlier at 11:28 UTC. The base model was originally published by Qwen on September 14.
According to Unsloth's documentation page, the GGUF variants can run on systems with 12-16 GB of VRAM at 1024x1024 resolution, while FP8 variants with CPU offloading can operate on as little as 6 GB of VRAM. The documentation notes that these configurations assume batch size 1 and single-image generation.
Vendor caveats on performance and format choice
Unsloth's guide explicitly states that 'Memory figures are estimates, not tested minimums' and that 'Memory usage varies with resolution and offloading settings.' The documentation further advises that 'FP8 quants are recommended for GPUs with 24 GB of VRAM or more' and may be preferable for any GPU setup because it 'can offer faster inference than GGUF models even with offloading.'
The GGUF format is described as primarily intended for 'CPU RAM, or use a unified-memory system such as a Mac.' Unsloth also published fidelity measurements comparing INT8 and FP8 quantizations against the base model, reporting INT8 LPIPS mean of 0.064 and SSIM mean of 0.936 versus FP8 LPIPS mean of 0.112 and SSIM mean of 0.899. These are vendor-published figures without independent third-party verification.
Research-only license constrains practical adoption
Both the base Qwen-Image-2.1 model and Unsloth's derivative quantizations carry the Qwen Research License Agreement, dated September 20, 2026, from Hangzhou Tongyi Laboratory Technology Co., Ltd. Section 2(a) grants rights 'FOR NON-COMMERCIAL PURPOSES ONLY,' defining non-commercial as 'for research or evaluation purposes only.'
Section 2(b) requires obtaining a separate commercial license for any commercial use. Section 4(b) mandates that derivative works prominently display 'Built with Qwen' or 'Improved using Qwen' in product documentation, while Section 4(c) prohibits using 'Qwen' as the primary name or identifier of derivative products. Section 8 designates the People's Courts in Hangzhou City as having exclusive jurisdiction over disputes.
Sources & context
Go to the original material. Company claims remain attributed to their sources.
01
