Qwen-Image 2.1

Model card checked 2026-09-26 · Memory fit estimated; speed unknown unless a separately linked measurement matches your configuration.

Also searched as: qwen image 2.1 / qwen-image-2.1 / qwen image 2.1 gguf

A text-to-image and image-editing pipeline, not an LLM: read the fit label as 'does not fit resident' rather than 'impossible', because the same pipeline is documented running on a 24 GB card with the text encoder offloaded. Qwen publishes no VRAM table and no quantization-tier mapping. The model card, the ModelScope card, the GitHub README and the release blog only offer enable_model_cpu_offload(); the official organization ships no quantized checkpoint, so every GGUF, INT8 and FP4 build is third-party. The 33.1 GB figure is the vLLM-Omni recipe's on-disk component total (Qwen3-VL 8B text encoder 17.5 GB + 7.1B DiT 14.2 GB + VAE 1.4 GB), not a measured peak. The two runtime projects that publish measured peaks disagree by design: SGLang's 2026-09-20 table reports 38.4 GiB resident on an H200 and a 22.7 GiB peak on a 24 GB RTX 4090 whose encoder is offloaded layer by layer, while vLLM-Omni reports 34.0 GB BF16 and 27.5-28.1 GB with text-encoder FP8 on a GB300 (NVIDIA only, and not yet in a tagged release). Community reports for the same 1024x1024 regime span 2 GB to 38.4 GiB with no shared protocol, so no single figure is 'the' requirement. GGUF also needs a loader that recognizes qwen_image 2.1 and a Qwen3-VL mmproj vision tower. See the memory and quantization decision page for every source and its conditions.