Unsloth Slashes Qwen-Image-2.1 Weight Storage in Half for Consumer GPUs

Unsloth has published FP8 and INT8 builds of Alibaba’s Qwen-Image-2.1, reducing the storage required for its largest components from 31.73 GB to about 16.5 GB. The smaller weights make local generation and editing more practical on consumer GPUs, particularly when paired with CPU offloading.

Quantization stores model weights with lower-precision numbers. FP8 and INT8 use 8 bits per value instead of the 16 bits used by BF16, cutting storage and weight memory while introducing some approximation.

At publication time, Unsloth’s model repository showed more than 850,000 downloads. Hugging Face counts file fetches rather than unique users, so the figure indicates distribution activity rather than adoption by 850,000 people.

Two components get the 8-bit treatment

The release provides quantized replacements for the pipeline’s two largest components: the diffusion transformer, or DiT, and the Qwen3-VL text encoder. Both use the safetensors format, which avoids executable pickle payloads during checkpoint loading.

The FP8 transformer and text encoder total 16.51 GB, compared with 31.73 GB for their BF16 equivalents. These figures cover checkpoint storage for the listed components. They exclude runtime activations, caches, other pipeline files, and framework overhead.

Quality trade-offs, in numbers

Unsloth compared same-seed outputs with the BF16 reference using LPIPS, a metric that measures perceptual differences between images. Lower scores indicate closer outputs. The company reports 0.064 for INT8 and 0.112 for FP8, which led it to select INT8 as the default in Unsloth Desktop.

LPIPS measures visual similarity rather than prompt adherence, text accuracy, or editing precision, so application-specific testing remains necessary. Unsloth also offers a Dynamic FP8 variant that reportedly runs with 6 GB of VRAM through CPU offloading, with a slowdown below 2x in its tests.