This release packages Qwen-Image-2.1 into a variety of quantized formats so you can run the model locally with clear size/performance tradeoffs. It’s aimed at users who want to run Qwen’s image model off-line (or on-device) and need pragmatic choices between VRAM use, inference speed, and image quality.
What Sets It Apart
- Multiple quantizations and formats: BF16, FP8, INT8 ConvRot, NVFP4, MLX (4/6/8-bit) and several GGUF variants (Q8_0, Q6_K, Q5_K_M, Q4_K_M, Q4_0), giving explicit options for different hardware and memory constraints. This means you can pick a quantization for maximum fidelity (BF16/FP8) or for minimal VRAM footprint (Q4_K_M / MLX-4bit).
- Companion files packaged: text encoders (BF16 and an Int8 ConvRot variant) and a BF16 VAE are provided to simplify ComfyUI workflows and reduce hunting for compatible assets. The repo also calls out recommended pairings (e.g., Q4_K_M for GGUF and qwen3vl_8b_int8_convrot for low-RAM text encoding).
- Practical guidance for local setups: highlights keeping the diffusion GGUF in GPU VRAM while offloading the text encoder to system RAM to save 9–17 GB of VRAM; notes Apple-Silicon–oriented MLX files and NVFP4/INT8 variants for supported NVIDIA workflows.
Who It's For and Tradeoffs
Great fit if you need an offline/localizable Qwen-Image-2.1 build and want clear quantization choices for specific hardware (NVIDIA desktop GPUs, low-VRAM machines, or Apple Silicon). Look elsewhere if you require a model with an upstream safety/filtering layer or an officially supported commercial license; this release is explicitly uncensored (no built-in safety checker) and is distributed under the Qwen Research License. Expect tradeoffs: smaller quantizations reduce VRAM and speed up inference but can slightly degrade fine detail and typography compared to BF16/FP8. The provided recommendations (Q4_K_M for GGUF, int8 text encoder in RAM) are pragmatic starting points for most consumer GPUs.
