| Licensed by NVIDIA Corporation under the NVIDIA Open Model License | |
| This repository redistributes a QUANTIZED derivative of the model weights of | |
| nvidia/Qwen-Image-Flash (https://huggingface.co/nvidia/Qwen-Image-Flash). Use of these | |
| weights is governed by the NVIDIA Open Model License Agreement: | |
| https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license | |
| Additional Information: Apache License 2.0 (see LICENSE), as provided by the upstream | |
| repository. | |
| Qwen-Image-Flash is a DMD2 distillation of Qwen/Qwen-Image | |
| (https://huggingface.co/Qwen/Qwen-Image). | |
| Changes relative to upstream: | |
| - transformer/model-int8.safetensors: the 20.43B DiT quantized to int8 (attention, | |
| feed-forward and modulation linears, group size 64). The in/out projections, timestep | |
| embedding and final norm are left at full precision. | |
| - text_encoder/model-int8.safetensors: the Qwen2.5-VL-7B LANGUAGE MODEL quantized to | |
| int8 (group size 64). The vision tower is omitted — text-to-image conditions on text | |
| alone and never invokes it. | |
| - vae/: unquantized, unchanged. | |
| - tokenizer/tokenizer.json: converted from the upstream slow-tokenizer files | |
| (vocab.json + merges.txt) and verified to produce identical token ids. | |