A complete Hugging Face pipeline bundle with the Qwen-Image-2.1 DiT in NF4 and a compact Qwen3.5-0.8B NF4 student text encoder. A learned MLP adapter maps Qwen3.5's 1024-dimensional token features into Qwen-Image's 4096-dimensional text-conditioning space. VAE, scheduler, Qwen3-VL-compatible pipeline processor, student tokenizer, adapter, inference helper, and a generated sample are included. The adapter was distilled from the pinned Qwen-Image Qwen3-VL teacher on 192 synthetic training prompts and 32 validation prompts. Validation token-embedding cosine similarity reached 0.9392. This is an embedding alignment metric, not a guarantee of image-quality parity; the included sample is an…
Open-weight model · Text to image
Qwen-Image-2.1-NVFP4
by Josep López joseplcam/Qwen-Image-2.1-NVFP4
Qwen-Image-2.1-NVFP4 is an open-weight model for text to image from Josep López, released under other. It has 4.5B parameters. At 16-bit it needs about 10.8 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.
Built with Qwen. Non-commercial research and evaluation use only (Qwen Research License, see LICENSE and NOTICE).
Parameters4.5B
Context—
Weights18.1 GB
Licenseother
AccessOpen weights
Monthly Downloads—
Runs On
What it takes to serve Qwen-Image-2.1-NVFP4 (4.5B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 9.0 GB | 10.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 4.5 GB | 5.4 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 2.3 GB | 2.7 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.
Qwen-Image-2.1-NVFP4 on every accelerator the SAVRN Index prices, at every precision
Model Card
Built with Qwen. Non-commercial research and evaluation use only (Qwen Research License, see LICENSE and NOTICE). An NVFP4 quantization of Qwen/Qwen-Image-2.1, made from the official BF16 weights with SVDQuant + GPTQ and calibrated on both text-to-image prompts and image edits. Both large components run on native Blackwell FP4 tensor cores, and the pipeline loads with a single DiffusionPipeline.frompretrained call. At 512x512 on an RTX PRO 6000 Blackwell it is about 1.7x faster than BF16 and needs about 40% less VRAM, and it matches BF16 closely on text-to-image, typography and edits, including a hard recolor edit that earlier 4-bit text encoders failed. trustremotecode=True loads two small…
Excerpt from the card by Josep López, licensed other.
Identity and Version
- Repository
- joseplcam/Qwen-Image-2.1-NVFP4
- Publisher
- Josep López
- Task
- Text to image
- Modality
- Image
- Library
- diffusers
- Parameters
- 4.5B parameters
- Languages
- Not stated by the source
- Revision
- c10a8f8405b64c4088bfffd0822dc16cec455fe5
- First published
- 2026-09-24
- Last updated
- 2026-09-25
Files and Weights
31 files, 18.1 GB in total. The weights are 3 files totalling 18.1 GB in safetensors.
Weights3 files · 18.1 GB
Configuration19 files · 54.3 KB
Tokenizer4 files · 15.9 MB
Documentation3 files · 18.1 KB
Other1 file · 5.3 KB
Repository1 file · 1.6 KB
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| text_encoder/model.safetensors | Weights | 11.8 GB | 83c4462b5ce6 |
| transformer/diffusion_pytorch_model.safetensors | Weights | 5.6 GB | 0304fc3e84b1 |
| vae/diffusion_pytorch_model.safetensors | Weights | 675.5 MB | 71879ffd5321 |
| model_index.json | Configuration | 497 B | — |
| processor/added_tokens.json | Configuration | 707 B | — |
| processor/preprocessor_config.json | Configuration | 782 B | — |
| processor/special_tokens_map.json | Configuration | 613 B | — |
| processor/video_preprocessor_config.json | Configuration | 817 B | — |
| scheduler/scheduler_config.json | Configuration | 485 B | — |
| text_encoder/config.json | Configuration | 6.4 KB | — |
| text_encoder/generation_config.json | Configuration | 213 B | — |
| text_encoder/modeling_nunchaku_qwen3vl.py | Configuration | 3.5 KB | — |
| tools/evaluate.py | Configuration | 4.6 KB | — |
| tools/modeling_nunchaku_qwen3vl.py | Configuration | 3.5 KB | — |
| tools/modeling_nunchaku_qwenimage21.py | Configuration | 1.6 KB | — |
| tools/package.py | Configuration | 4.0 KB | — |
| tools/parrot_test.py | Configuration | 2.9 KB | — |
| tools/quantize_dit.py | Configuration | 5.3 KB | — |
| tools/quantize_text_encoder.py | Configuration | 5.2 KB | — |
| transformer/config.json | Configuration | 9.5 KB | — |
| transformer/modeling_nunchaku_qwenimage21.py | Configuration | 1.6 KB | — |
| vae/config.json | Configuration | 2.1 KB | — |
| LICENSE | Documentation | 7.8 KB | — |
| NOTICE | Documentation | 1.8 KB | — |
| README.md | Documentation | 8.5 KB | — |
| processor/chat_template.jinja | Other | 5.3 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| processor/merges.txt | Tokenizer | 1.7 MB | — |
| processor/tokenizer.json | Tokenizer | 11.4 MB | aeb13307a71a |
| processor/tokenizer_config.json | Tokenizer | 5.4 KB | — |
| processor/vocab.json | Tokenizer | 2.8 MB | — |
License and Download
- License
- other
- Access
- Open weights, no gate
- Download size
- 18.1 GB
Download from Josep López
Released by Josep López through its official repository on Hugging Face.
Built From
- Derived from Qwen/Qwen-Image-2.1
- Described by arXiv:2411.05007
- Quantized from Qwen/Qwen-Image-2.1
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 18.1 GB |
| 16-bit | 9.0 GB |
| 8-bit | 4.5 GB |
| 4-bit | 2.3 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About Qwen-Image-2.1-NVFP4
How much GPU memory does Qwen-Image-2.1-NVFP4 need?
About 10.8 GB at 16-bit and 2.7 GB at 4-bit: the weights (4.5B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run Qwen-Image-2.1-NVFP4 on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
What license is Qwen-Image-2.1-NVFP4 released under?
other, as its publisher declares it. Read the license text before commercial use.
Similar Models
Welcome to the official repository for the Z-Image（造相）project! Z-Image is a powerful and highly efficient image generation model family with 6B parameters. Currently there are four variants: - Z-Image-Turbo – A distilled version of Z-Image that matches or exceeds leading competitors with only 8 NFEs (Number of Function Evaluations). It offers sub-second inference latency on enterprise-grade H800 GPUs and fits comfortably within 16G VRAM consumer devices. It excels in photorealistic image generation, bilingual text rendering (English & Chinese), and robust instruction adherence. - Z-Image – The foundation model behind Z-Image-Turbo. Z-Image focuses on high-quality generation, rich…
basemodel: stabilityai/stable-diffusion-xl-base-1.0 - stable-diffusion-xl - stable-diffusion-xl-diffusers - text-to-image - diffusers - inpainting SD-XL Inpainting 0.1 is a latent text-to-image diffusion model capable of generating photo-realistic images given any text input, with the extra capability of inpainting the pictures by using a mask. The SD-XL Inpainting 0.1 was initialized with the stable-diffusion-xl-base-1.0 weights. The model is trained for 40k steps at resolution 1024x1024 and 5% dropping of the text-conditioning to improve classifier-free classifier-free guidance sampling. For inpainting, the UNet has 5 additional input channels (4 for the encoded masked-image and 1 for the…
SDXL consists of an ensemble of experts pipeline for latent diffusion: In a first step, the base model is used to generate (noisy) latents, which are then further processed with a refinement model (available here: https://huggingface.co/stabilityai/stable-diffusion-xl-refiner-1.0/) specialized for the final denoising steps. Note that the base model can be used as a standalone module. Alternatively, we can use a two-stage pipeline as follows: First, the base model is used to generate latents of the desired output size. In the second step, we use a specialized high-resolution model and apply a technique called SDEdit (https://arxiv.org/abs/2108.01073, also known as "img2img") to the latents…
SDXL-Turbo is a fast generative text-to-image model that can synthesize photorealistic images from a text prompt in a single network evaluation. A real-time demo is available here: http://clipdrop.co/stable-diffusion-turbo Please note: For commercial use, please refer to https://stability.ai/license. SDXL-Turbo is a distilled version of SDXL 1.0, trained for real-time synthesis. SDXL-Turbo is based on a novel training method called Adversarial Diffusion Distillation (ADD) (see the technical report), which allows sampling large-scale foundational image diffusion models in 1 to 4 steps at high image quality. This approach uses score distillation to leverage large-scale off-the-shelf image…
Animagine XL 4.0, also stylized as Anim4gine, is the ultimate anime-themed finetuned SDXL model and the latest installment of Animagine XL series. Despite being a continuation, the model was retrained from Stable Diffusion XL 1.0 with a massive dataset of 8.4M diverse anime-style images from various sources with the knowledge cut-off of January 7th 2025 and finetuned for approximately 2650 GPU hours. Similar to the previous version, this model was trained using tag ordering method for the identity and style training. With the release of Animagine XL 4.0 Opt (Optimized), the model has been further refined with an additional dataset, improving stability, anatomy accuracy, noise reduction…