Qwen-Image server: RTX PRO 4500 or 6000 Blackwell Server Edition, VRAM and images per hour

Qwen-Image GPU requirements: which card holds it

Qwen-Image runs in FP8 on one NVIDIA RTX PRO 4500 Blackwell Server Edition with 32 GB and in full BF16 on one NVIDIA RTX PRO 6000 Blackwell Server Edition with 96 GB. Its 20B diffusion transformer takes 40.9 GB in BF16, which no 32 GB card holds, and about 20.4 GB as an FP8 file. For product images at volume, eight RTX PRO 4500 Server Edition cards and four RTX PRO 6000 Server Edition cards deliver a similar number of images per kilowatt by our estimate; the choice depends on precision, editing workflows and what else the server runs.

Alibaba's Qwen team released Qwen-Image with its weights on 4 August 2025 under Apache 2.0, followed by Qwen-Image-Edit, Edit-2509, Edit-2511, Qwen-Image-Layered and Qwen-Image-2512, whose repositories were last updated between August and December 2025. Our FLUX GPU requirements guide lists Qwen-Image next to FLUX and Stable Diffusion with their licences; this article sizes the Apache-licensed Qwen-Image models on the two Server Edition cards, for e-commerce teams and agencies that generate and edit product imagery every day.

NVIDIA product pages and product briefs, read 24 October 2026; FP8 rates are StorageReview's MAMF measurements of sustained dense matrix multiplication, 14 October 2026, the RTX PRO 6000 figure on the Workstation Edition, which StorageReview says profiles identically to the Server Edition.

Qwen-Image memory: BF16, FP8 and NVFP4 files

The pipeline has three parts: the diffusion transformer, a text encoder and a VAE. The text encoder is Qwen2.5-VL with 28 layers and a hidden size of 3,584, the 7B vision-language model, as its configuration and the file names of the ComfyUI repackaging show. The official repository is 57.7 GB.

Hugging Face file lists of Qwen/Qwen-Image (BF16) and Comfy-Org/Qwen-Image_ComfyUI (fp8_e4m3fn, fp8_scaled and nvfp4 files), read 24 October 2026; VAE size our arithmetic as the difference between the repository total and the two large folders.

On the RTX PRO 6000 Server Edition, whose driver reports 95.6 GiB, the BF16 pipeline leaves about 42 GiB for activations, a second model or a LoRA set. On the RTX PRO 4500 Server Edition we estimate about 31.9 GiB visible, scaled from the 96 GB card, so the FP8 pipeline leaves about 4 GiB. Activations at the card's resolutions are small by our arithmetic (one hidden state of 6,889 tokens × 3,072 values in BF16 is about 42 MB), but the room is too small for a second model. Diffusers and ComfyUI can also unload the text encoder after the prompt is encoded, which frees about 9 GB on the 4500 at the cost of reloading it for the next prompt. Our users-per-card rule for LLMs (90 per cent of the driver figure less 3 GiB) does not apply here, since a diffusion model keeps no per-user cache; we compare file sizes with the memory the driver reports and leave headroom.

The 32 GB card does not hold the BF16 transformer even on its own. FP8 halves the weights, and quality has to be checked on your own prompts and products before you size on it. The ComfyUI repository also lists an "fp8_hq" variant of 22.7 GB and a mixed FP8 file of 20.5 GB. NVFP4 is the 4-bit format that the FP4 Tensor Cores of Blackwell compute, and both RTX PRO cards have them; we found no published speed or quality comparison for them, and StorageReview notes for the 4500 that "the FP4 advantage exists at the GEMM level and has not reached the serving stack yet" for LLMs.

Images per hour: what one image costs

The Qwen-Image card's example generates at 1328 × 1328 for square images and 1664 × 928 for 16:9, with num_inference_steps=50 and true_cfg_scale=4.0. Diffusers' documentation states that classifier-free guidance "is enabled by setting true_cfg_scale > 1 and a provided negative_prompt", even an empty one, and then the transformer runs twice per step. The VAE compresses each side by 8 and the transformer packs 2 × 2 latent pixels into a token, so one token covers 16 × 16 pixels by our reading of the configurations: 6,889 image tokens at 1328 × 1328, 6,032 at 1664 × 928.

Qwen-Image's transformer has two streams. In each of its 60 blocks the image tokens and the text tokens have their own attention projections and feed-forward layers and meet only in the joint attention, and the modulation layers act once per image on the timestep embedding, as the Diffusers code shows. An image token therefore passes through about 6.8B of the 20.4B weights, 4 × 3,072² + 2 × 3,072 × 12,288 per block. By the standard count a pass costs about 2 × 6.8B × 6,889 operations in the linear layers plus 4 × 6,889² × 3,072 × 60 in attention, 9.4 × 10^13 plus 3.5 × 10^13, about 1.3 × 10^14 together. The prompt tokens, the text encoder and the VAE come on top and are left out, which keeps the figure a floor. One hundred passes, 50 steps with guidance, come to about 1.3 × 10^16 operations per image. The Lightning LoRA from lightx2v, under Apache 2.0, runs Qwen-Image in 8 steps with true_cfg_scale=1.0, as both its card and the Diffusers documentation show: 8 passes instead of 100.

Dividing by StorageReview's measured FP8 matrix rate gives a floor for the time per image and an upper bound for images per hour. Real pipelines reach less than a pure matrix benchmark, and a BF16 run computes more slowly than FP8, so measured rates will be lower. Whether an FP8 file computes in FP8 or only stores its weights in FP8 depends on the software; the bound holds either way. On the RTX PRO 6000, a BF16 run is bounded more tightly by NVIDIA's dense BF16 peak of 503.8 TFLOPS (Workstation Edition; the Server Edition is rated lower), at most about 141 images per hour at 50 steps. An NVFP4 compute path is not covered: StorageReview measured 569 and 1,481 TFLOPS in FP4 on the two cards, and we found no published Qwen-Image speed figure for it.

Our estimate of upper bounds at 1328 × 1328, one job per card, for FP8 or BF16 computation, rounded down: 1.29 × 10^16 operations per image at 50 steps with guidance and 1.03 × 10^15 with Lightning at 8 steps, from the model card, the configuration files and the Diffusers code, divided by StorageReview's FP8 MAMF results of 14 October 2026 (280 TFLOPS at 165 W, 774 TFLOPS).

We found no published images per second for Qwen-Image on either card. NVIDIA's inference performance page (metadata 19 October 2026) lists FLUX, Stable Diffusion XL and Stable Video Diffusion for the RTX PRO 6000 Server Edition but no Qwen-Image row. The Diffusers documentation shows a torch.compile speed-up of about 2.4 times on an A100 80GB without stating resolution or guidance, so we do not use its absolute times.

For a shop that refreshes 10,000 product images, the table gives a lower bound of about 1.3 hours with Lightning on eight RTX PRO 4500 Server Edition cards and about 16 hours at 50 steps, by our arithmetic. Measure your own prompts at your step count and resolution on one card, then divide the images of your busiest hour by that rate. We supply both cards and build servers around them; send us your images per day, resolutions and edit workflows, and our team replies with a configuration for each card.

Editing workflows with Qwen-Image-Edit-2511

Qwen-Image-Edit-2511 is a 20B model in BF16 under Apache 2.0. Its card lists better character consistency, "mitigate image drift", the fusion of two separate person images into one coherent picture and "integrated LoRA capabilities", and its example passes two input images to the QwenImageEditPlusPipeline with 40 steps and true_cfg_scale 4.0. An edit also processes the input pictures, so it costs more per image than a generation at the same size; we found no published figure and recommend measuring your own edits.

Our article on the AI server for e-commerce puts image editing next to product texts, search and a shop chatbot on one server, and our AI server for creative agencies next to video and rendering. Both use the RTX PRO 6000 Server Edition, because one 96 GB card can hold a 20B edit model in BF16, about 41 GB at 2 bytes per parameter, next to the 20.4 GB FP8 generator.

Eight RTX PRO 4500 or four RTX PRO 6000 Server Edition

Our estimates from the tables above; power as on our configuration pages (cards at 165 or 600 W, two AMD EPYC processors at 280 W, plus 15 per cent).

The eight-card RTX PRO 4500 server gives each workflow its own card and stays on one 230 V 16 A feed per side by our estimate. Its 16 GB MIG instances are too small for the 20.4 GB FP8 transformer, so on this card Qwen-Image runs on whole cards. Our page on the eight-card RTX PRO 4500 Server Edition server covers chassis and power, and the four-card 2U build the smaller variant at about 1.4 kW.

The four-card RTX PRO 6000 server runs Qwen-Image in BF16 when quality matters more than throughput, holds an edit model and a generator on the same card, and can give a card to an LLM for product texts. Split by MIG into two 48 GB instances, one card carries two FP8 pipelines of about 30 GB each, or four 24 GB instances that hold the FP8 transformer with the encoder unloaded. Our comparison of eight RTX PRO 6000 or eight RTX PRO 4500 Server Edition weighs the two cards for LLM work, and several RTX PRO 5000 or 4500 against fewer RTX PRO 6000 gives the general trade-off.

A rule of thumb from these figures: choose the RTX PRO 4500 Server Edition when FP8 quality is accepted after a test, many people submit at once and the room has only 16 A feeds; choose the RTX PRO 6000 Server Edition for BF16, for edits with several input images, or when the same server also runs LLMs or video. We size servers with either card and check rack, feed and airflow before we quote; describe the models you plan and the people generating at peak.

Licences and text in images

The Qwen-Image card names Apache 2.0, and the repository's LICENSE file is the Apache License, Version 2.0, with no added territory or commercial clauses. The cards of Qwen-Image-2512, Qwen-Image-Edit-2511 and the Lightning LoRA also name Apache 2.0, as does the ComfyUI repackaging. Qwen's newer Qwen-Image-2.1 is under a non-commercial research licence, as our FLUX guide sets out, so check the licence of every checkpoint you put into production; whether a specific use is covered is a question for your legal department.

The card describes "exceptional performance in text rendering, especially for Chinese", and names "alphabetic languages like English". It makes no claim for German, French or Polish lettering with diacritics, so test product labels in your languages. Under Article 50 of the EU AI Act, providers of generative systems mark generated images in a machine-readable format, as our article on the EU AI Act for companies explains as general information.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We supply the RTX PRO 4500 Server Edition and the RTX PRO 6000 Server Edition as cards for your servers or in AI servers built to order, assembled and tested before delivery, on one EU contract and invoice with manufacturer warranty on every component. We install the operating system, drivers, CUDA and a container runtime on request; quantities and lead time are confirmed in writing with the quote.

FAQ