ComfyUI on NVIDIA RTX PRO 2000, 4000 and 4500 Blackwell: which SD 3.5 and FLUX files fit

Which image models fit 16, 24 and 32 GB in ComfyUI

On the 16 GB NVIDIA RTX PRO 2000 Blackwell, ComfyUI can keep SDXL and Stable Diffusion 3.5 Medium with all their parts in GPU memory. The 24 GB NVIDIA RTX PRO 4000 Blackwell does the same for FLUX.2 klein 4B and, with the FP8 text encoder, for Stable Diffusion 3.5 Large. The 32 GB NVIDIA RTX PRO 4500 Blackwell Workstation Edition adds Stable Diffusion 3.5 Large with its 16-bit text encoder and the model file of FLUX.1-dev in 16-bit, which a company may only use with a commercial licence.

These statements are arithmetic. We took the sizes of the publishers' own files on Hugging Face on 10 November 2026 and set them against 90 per cent of each card's memory. Nothing was run, and we found no published image-generation measurement for these three cards, so this article contains no seconds per image.

How the fit is counted

A ComfyUI workflow for a current image model loads three kinds of files: the diffusion model, one or more text encoders and a small VAE that turns the result into pixels. The single-file checkpoints of Stable Diffusion 3.5 and FLUX contain the diffusion model only. The text encoder comes separately and is often the part that does not fit beside the model: the T5-XXL encoder of Stable Diffusion 3.5 is a 9.79 GB file in 16-bit and 4.89 GB in FP8, and the text encoder of FLUX.2-dev alone has 48.02 GB.

Against these files we set 0.9 times the card's memory, counted on the nominal size: 14.4 GiB for 16 GB, 21.6 GiB for 24 GB, 28.8 GiB for 32 GB and 43.2 GiB for 48 GB. The driver on the delivered card reports slightly less than the nominal size, and that reading decides. The remaining tenth is left for what a run needs beyond the files; we have no published figure for that share, so a fit here means that the files fit, and says nothing about a particular resolution.

Three outcomes follow, and the table uses these words for them:

• complete: diffusion model, text encoders and VAE fit together;
• model only: the diffusion model fits, and the text encoder has to be unloaded or kept in system memory;
• partial: the model file alone exceeds the 90 per cent, so it runs only with partial loading from system memory, which is slower.

File sizes and licences of the models

file listings and licence tags of the publishers' repositories on Hugging Face (stabilityai, black-forest-labs, Qwen), read 10 November 2026; sums and conversion to GiB are our arithmetic.

our arithmetic from the first table against 14.4, 21.6, 28.8 and 43.2 GiB. The cards are the RTX PRO 2000 Blackwell (16 GB), RTX PRO 4000 Blackwell (24 GB), RTX PRO 4500 Blackwell Workstation Edition (32 GB) and RTX PRO 5000 Blackwell (48 GB). Not measurements.

Black Forest Labs also publishes FP8 and NVFP4 versions of FLUX.2 klein 4B. We did not read their sizes, so they are missing from the table, as is the newer FLUX.3 image klein family of October 2026.

You can order the workstation alone, the setup of ComfyUI with SDXL and FLUX.2 klein 4B alone, or both together: the workstation delivered with everything installed and configured; write to us and we suggest the best solution for your tasks and budget.

RTX PRO 2000 Blackwell: ComfyUI on 16 GB and 70 W

The RTX PRO 2000 Blackwell is the card for a compact office workstation: 16 GB of GDDR7 with ECC at 288 GB/s, 70 W without a power cable, half-height (2.7 in), the format usually called low profile, and two slots wide. SDXL fits completely with 6.46 GiB, and Stable Diffusion 3.5 Medium fits completely when the T5 encoder is the FP8 file. With the 16-bit encoder the Medium pipeline comes to 15.40 GiB, above the 14.4 GiB, and the encoder moves to system memory.

For FLUX.2 klein 4B the model file fits and the pipeline of 14.87 GiB misses the threshold by half a GiB, so the text encoder has to be unloaded or kept in system memory. Stable Diffusion 3.5 Large, at 15.33 GiB for the model alone, runs only with partial loading. Card pages: RTX PRO 2000 Blackwell specs, buy the RTX PRO 2000 Blackwell.

The NVIDIA RTX 2000 Ada Generation has the same 16 GB, on GDDR6 at 224 GB/s, and the same 70 W. The file arithmetic is identical for FP8 and 16-bit files. The difference lies in NVFP4: NVIDIA documents FP4 support for the fifth-generation Tensor Cores of the Blackwell cards, while the Ada cards have FP8 and no FP4 acceleration in hardware. On an RTX 2000 Ada, choose FP8 or 16-bit files. Card pages: RTX 2000 Ada specs, buy the RTX 2000 Ada. What the same 16 GB hold as a language model is the subject of our article on local LLMs on 16 GB.

RTX PRO 4000 Blackwell: FLUX.2 klein and SD 3.5 Large on 24 GB

With 24 GB two more models fit completely. FLUX.2 klein 4B, the Apache 2.0 model of the FLUX.2 family, fits with all its parts at 14.87 GiB. Stable Diffusion 3.5 Large fits completely with the FP8 text encoder, at 21.41 GiB against 21.6 GiB. That margin of 0.19 GiB is counted on the nominal size; if the driver reports 23.9 GiB, as the scaling from the 96 GB card on our language-model pages assumes, the threshold is 21.5 GiB and the margin 0.1 GiB. With the 16-bit encoder it is a model-only case. FLUX.1-dev in NVFP4 would fit completely as well, and its licence excludes commercial use.

The RTX PRO 4000 Blackwell has 24 GB of GDDR7 with ECC at 672 GB/s, 8,960 CUDA cores and 145 W in a single full-height slot, fed through one 16-pin connector. NVIDIA lists neither MIG nor vGPU for this card. Because it takes one slot, several cards fit one tower, as our page on four RTX PRO 4000 in one workstation describes. Card pages: RTX PRO 4000 Blackwell specs, buy the RTX PRO 4000 Blackwell; the step from 16 GB is compared in RTX PRO 2000 vs RTX PRO 4000 Blackwell, and language models on this card in local LLMs on the RTX PRO 4000 Blackwell.

RTX PRO 4500 Blackwell Workstation Edition: 32 GB for the 16-bit files

The 32 GB card takes the files that are too large for 24 GB. Stable Diffusion 3.5 Large fits completely with the 16-bit T5 encoder (25.97 GiB against 28.8 GiB), and FLUX.2 klein 4B leaves about 14 GiB free by the same count. The 16-bit model file of FLUX.1-dev fits with 22.17 GiB while its whole pipeline of 31.58 GiB does not, and FLUX.2-dev in NVFP4 remains a model-only case on the 24, 32 and 48 GB cards, because its text encoder is larger than the model. Both are non-commercial models.

The RTX PRO 4500 Blackwell Workstation Edition has 32 GB of GDDR7 with ECC at 896 GB/s, 10,496 CUDA cores and 200 W in two slots; an adapter from two 8-pin PCIe cables is in the box by NVIDIA's quick start guide. It is a different card from the passively cooled RTX PRO 4500 Server Edition for rack servers; our comparison of the RTX PRO 4500 and the RTX PRO 4500 Server Edition explains the two. Card pages: RTX PRO 4500 Blackwell 32 GB, buy the RTX PRO 4500 Blackwell, and the comparison RTX PRO 4000 vs RTX PRO 4500 Blackwell.

The next step is the NVIDIA RTX PRO 5000 Blackwell with 48 GB, the first card in the table on which the Apache 2.0 model Qwen-Image fits as a 16-bit model file and FLUX.1-dev in 16-bit fits completely: RTX PRO 5000 with 48 or 72 GB, buy the RTX PRO 5000 Blackwell 48 GB. Larger models and cards are covered in our guide to FLUX GPU requirements.

Licences: which models a company may use

For a company the first filter is the licence tag of the repository, and only then the size.

• Apache 2.0: FLUX.2 klein 4B and Qwen-Image. Of FLUX.2 klein only the 4B variant carries this tag.
• OpenRAIL++: SDXL base 1.0, with commercial use permitted within the use restrictions of the licence.
• Stability AI Community License: Stable Diffusion 3.5 Medium and Large. Commercial use is tied to conditions in the licence file, which we did not re-read for this article and do not summarise; both repositories are gated.
• Non-commercial licences: FLUX.1-dev and FLUX.2-dev in every format, including the NVFP4 files. Using them in a company's production needs a commercial licence from Black Forest Labs.

Whether a given use counts as commercial under one of these licences is a question for your legal counsel. Our guide to FLUX GPU requirements sets the licences of these model families side by side.

NVFP4, FP8 and what the AI TOPS figure means

NVFP4 files are far smaller than the 16-bit files in the table: 8.56 against 22.17 GiB for FLUX.1-dev, 19.59 against 60.02 GiB for FLUX.2-dev. They compute natively only on the Blackwell cards, whose compute capability NVIDIA gives as 12.0; the Ada cards are listed with 8.9, which covers FP8. NVIDIA's support matrix for its visual generative AI microservices of 20 October 2026 ties the formats to that figure: bf16 needs compute capability 8.0, fp8 needs 8.9 and nvfp4 needs 10.0.

NVIDIA's datasheets rate the RTX PRO 2000 Blackwell at 545 AI TOPS and the RTX PRO 4500 Blackwell Workstation Edition at 1,617; for the RTX PRO 4000 Blackwell, NVIDIA's datasheet (June 2026) gives 1,178 AI TOPS and its product page 1,290. The figure is a theoretical FP4 rate with sparsity and no measured speed, so it does not translate into images per minute.

What ComfyUI does when a model does not fit

ComfyUI's README, read on 10 November 2026, describes "smart VRAM and RAM management, model offloading, and support for quantized models" and says the software can run even the biggest open source models "on as low as 4GB vram + 8GB ram relatively quickly", using what it calls asynchronous weight streaming. Its list of supported image models includes SDXL, SD3.5, Flux.1, Flux.2 and Qwen Image. The README we read describes no command-line switch for low-memory operation, so we name none.

Offloading moves the load to system memory. In a model-only case the text encoder waits there: 9.79 GB for the 16-bit T5 encoder, and about 45 GiB for the encoder and VAE of FLUX.2-dev by the difference of the two table columns. NVIDIA's quick start guide for the Blackwell cards asks for at least as much system memory as GPU memory and recommends twice as much, which is 32, 48 and 64 GB for the three cards. If you know the models, send us the list of models, the formats you want to run and the number of people who generate images, and we size the card and the system memory together.

NVIDIA's memory figures for the same models as NIM microservices

NVIDIA publishes memory requirements for the same models as NIM microservices, and they are a different count from the file arithmetic above. The support matrix of 20 October 2026 gives 16 GB as the minimum and 32 GB as recommended for FLUX.1-dev and for Stable Diffusion 3.5 Large, 48 GB for FLUX.2 klein and 80 GB for Qwen-Image. Its pre-built engines are listed for a set of named GPUs, and none of the cards in this article is among them. A team that plans to serve these models as NIM containers should size by NVIDIA's matrix; the tables above describe ComfyUI with the publishers' files.

On speed there is one statement from NVIDIA: its blog of 11 August 2025 speaks of a "1.4x boost in image generation" for the RTX PRO 2000 Blackwell compared with the previous generation, without naming the baseline card and without publishing test conditions. Measurements of the Ada cards in other AI tasks are collected in our article on AI on the RTX 5000 Ada and RTX 4500 Ada, and what the 4 and 8 GB cards can do is in AI on the RTX A1000 and RTX A400. Teams that move from stills to clips find the larger workstation in our article on the AI video workstation for marketing.

FAQ