FLUX 3 Klein hardware requirements: VRAM, GPU and ComfyUI
- Page updated
- Sources checked
- File sizes from Hugging Face, read October 17, 2026
- Memory and speeds as Black Forest Labs publishes them (RTX speeds measured by NVIDIA), with their setup
- Jump to the checker ↓
FLUX 3 Image Klein is Black Forest Labs' open-weight image model, released on October 15, 2026: a 5.6B-parameter image generator for text-to-image, editing with up to 10 reference images and layout with bounding boxes, in a 16-step distilled version and a 50-step base version for fine-tuning, under the FLUX Kommunity License. It is meant to run with a 9B prompt upsampler in front of it.
Short answer: Black Forest Labs says the int8 files in ComfyUI run on 8 GB GPUs such as the RTX 3070. BFL does not say whether its 8 GB figure includes the 9B upsampler it says not to skip, so by our reading on 8 GB expect to run it separately or switch it off (with weaker results). BFL's own code has no int8 path, and its NVFP4 path needs a Blackwell GPU, which by our reading includes the RTX 50 series.
Match your GPU to Black Forest Labs' FLUX 3 Klein guidance
Open in a full page ↗Loading private checker…
- Developer
- Black Forest Labs Official
- Params
- 5.6B (image generator; text encoder not counted) Official
- Tasks
- Text-to-image, editing with up to 10 reference images, layout with bounding boxes, up to 4K Official
- License
- FLUX Kommunity License Official
- Weights
- Hugging Face, gated (accept the license first) Official
- ComfyUI
- the int8 files in ComfyUI run on 8 GB GPUs such as the RTX 3070 Official
- BFL reference code
- BFL's own code has no int8 path Official
- Upsampler
- BFL does not say whether its 8 GB figure includes the 9B upsampler Official
How much VRAM does FLUX 3 Klein need?
File sizes of the ComfyUI files Black Forest Labs publishes (flux-3-image-klein-comfy, decimal GB, read October 17, 2026). In the file names, base marks the base model; without it, the step-distilled model. BFL renamed these files on October 17, 2026 (sizes unchanged).
| Component | File | Size | Needs | Source |
|---|---|---|---|---|
| Image model, distilled, BF16 | flux3-image-klein-bf16 | 11.10 GB | room for 11.1 GB plus the encoder (our reading) | BFL |
| Image model, distilled, int8 | flux3-image-klein-int8-convrot | 5.83 GB | ComfyUI (the 8 GB route) | BFL |
| Image model, distilled, NVFP4 | flux3-image-klein-nvfp4 | 5.28 GB | Blackwell (BFL, for its own code); RTX 50 by our reading | BFL |
| Image model, base, BF16 / int8 / NVFP4 | flux3-image-klein-base-bf16 … | 11.08 / 5.81 / 5.26 GB | as above | BFL |
| Text encoder (Qwen3-VL 4B), BF16 | qwen3vl_4b_bf16 | 8.88 GB | one of the two required | BFL |
| Text encoder, int8 | qwen3vl_4b_int8_convrot | 4.86 GB | one of the two required | BFL |
| Prompt upsampler (9B), BF16 | qwen3.5_9b_flux3_prompt_upsampler_bf16 | 18.82 GB | BFL: do not skip (one of the three); added to this repository on October 17, 2026 | BFL |
| Prompt upsampler, int8 | qwen3.5_9b_flux3_prompt_upsampler_int8_convrot | 9.75 GB | BFL: do not skip (one of the three) | BFL |
| Prompt upsampler, w4a8 | qwen3.5_9b_flux3_prompt_upsampler_w4a8 | 6.72 GB | BFL: do not skip (one of the three) | BFL |
| VAE | flux2-vae | 0.34 GB | required | BFL |
The Diffusers-format repositories (flux-3-image-klein and -base) carry the same image model in BF16 (11.1 GB) plus FP8 (6.27 / 6.25 GB) and NVFP4 builds under variants/. BFL's blog quotes the image model at 11.2 GB in BF16 and 5.6 GB in FP8; the files are 11.10 and 6.27 GB. The upsampler in BF16 (flux-3-image-klein-pup) is 18.8 GB. BFL's download of the distilled model with its upsampler for its own code is 52 GB; everything (both models) is 107 GB. Community GGUF files of the image model exist (Abiray 1.99–5.99 GB, realrebelai 3.37–6.07 GB); we have not checked whether the ComfyUI-GGUF node loads FLUX 3, and BFL does not publish or test them.
What GPU do you need for FLUX 3 Klein?
| GPU | What the sources say | Source |
|---|---|---|
| 8 GB (RTX 3070 class) | BFL recommends the int8 files in ComfyUI for an 8 GB GPU (RTX 3070 class): they "run on 8 GB GPUs such as the RTX 3070". The blog: "as little as 8GB of VRAM, through ComfyUI or Diffusers". BFL does not say whether this includes the prompt upsampler | BFL GitHub |
| RTX 30 series and older | No FP8: use the int8 ComfyUI files (BFL's own code has no int8 path) | BFL GitHub |
| RTX 40 (Ada), Hopper | BFL's code adds --fp8; ComfyUI int8 is BFL's RTX benchmark build | BFL GitHub |
| RTX 50 and other Blackwell | --nvfp4 in BFL's code (not with --offload), or the NVFP4 ComfyUI files | BFL GitHub |
| B200 (BFL's test GPU) | Peak 31 GB (BF16), 26 GB (FP8), 25 GB (NVFP4) for the distilled model at 1024×1024 with BFL's CLI, upsampler off (nvidia-smi; text-to-image and a one-reference edit). With the upsampler on, it reserves 35% of GPU memory by default | BFL GitHub |
On an 8 GB card the int8 image model (5.83 GB), int8 text encoder (4.86 GB) and VAE (0.34 GB) add up to 11.0 GB, more than the card holds, so by our reading ComfyUI keeps them on the GPU one at a time (it runs the text encoder first, then the image model) and needs system RAM for the rest; BFL gives no RAM figure. BFL's peak-memory table is for its own reference code on a B200 and does not apply to ComfyUI's memory management. We found no published figure for Apple silicon.
Do you need the prompt upsampler?
Black Forest Labs says not to use either model without it: the model is trained on the captions the upsampler writes, and a prompt sent as typed gives "noticeably weaker results". It is a 9B vision-language model fine-tuned from Qwen3.5-9B (Apache 2.0), 18.8 GB in BF16, or 9.75 GB (int8) and 6.72 GB (w4a8) in the ComfyUI repository, whose text-to-image templates have an enable_prompt_upsampler switch. When the upsampler and the image model do not fit on the GPU together, BFL's README says to run the upsampler on its own first (python -m flux3.pup) and pass its caption with --no-pup; any other LLM can do the job with BFL's published system prompts.
Which tools run FLUX 3 Klein locally?
| Tool | Status (checked October 16, 2026) | Source |
|---|---|---|
| ComfyUI | Day-one support; ComfyUI master has a FLUX 3 model. BFL's flux-3-image-klein-comfy repository has BF16, int8 and NVFP4 files for both models, plus templates | BFL, ComfyUI |
| BFL reference code | CLI, Python API and a Gradio demo; BF16, FP8 (Ada / Hopper / Blackwell) and NVFP4 (Blackwell); an --offload option; tested on B200 | BFL GitHub |
| Diffusers | BFL lists support "from launch day"; Diffusers' main branch has FLUX 3 as a Modular Diffusers pipeline (base and distilled blocks), and the Hugging Face repositories ship a modular index. We have not checked which release includes it | BFL, Diffusers |
| SGLang Diffusion, vLLM-Omni, TensorRT-LLM | Day-one support; BFL names them as its serving frameworks | BFL |
| ComfyUI-GGUF | Community GGUF files of the image model exist; not published or tested by BFL | Hugging Face |
How fast is FLUX 3 Klein?
| Hardware | Setup | Time per image | Source |
|---|---|---|---|
| RTX 5090 | Distilled, int8, 16 steps, 1 MP, ComfyUI | about 4 s | BFL (measured by NVIDIA) |
| RTX 4070 | Distilled, int8, 16 steps, 1024×1024, ComfyUI | 8.9 s | BFL (measured by NVIDIA) |
| H200 | 1 MP, distilled (16 steps) vs base (50 steps with CFG); precision and software not stated | about 3 s vs 12 s | BFL GitHub |
| GB300 | Distilled, NVFP4, 16 steps, 1024×1024 | 0.55 s | BFL (measured by NVIDIA) |
Generation time on the GPU only, without the prompt upsampler. We found no published time for an 8 GB card; we do not convert these figures to other GPUs.
Measured FLUX 3 Image Klein on your own hardware, or tried and it failed? Send us the numbers: paste the benchmark output or log into the form (or email it). After checking it we add it to this page, labeled as a community report and shared under CC BY 4.0.
FLUX 3 Image Klein specs
| Parameters | 5.6B (image generator; text encoder not counted) BFL |
|---|---|
| Models | Distilled: 16 steps, no guidance. Base: 50 steps, CFG 4 (3.5 until BFL changed it on October 17, 2026), for fine-tuning and LoRA training BFL GitHub |
| Tasks | Text-to-image, editing with up to 10 reference images, layout with bounding boxes, up to 4K BFL |
| Text encoder | Qwen3-VL 4B (from the ComfyUI file name qwen3vl_4b_*) |
| Weights | Hugging Face, gated (accept the license first); BF16, FP8, int8 (ComfyUI) and NVFP4 Hugging Face |
| License | FLUX Kommunity License: free commercial use for individuals and companies under $5M annual revenue, plus research and non-commercial use; the upsampler is Apache 2.0 BFL |
Common setups
The checker's own answer for 8 common setups (graphics cards with 32 or 64 GB of system RAM, then Macs), in the finder's quick mode (Black Forest Labs' int8 ComfyUI route on the card). Answers are published route matches, not measured setups; "Unknown" means the checker has no route or source for that hardware. Your own setup: the checker.
| Setup | Answer |
|---|---|
| GeForce RTX 3060 (12 GB) + 32 GB RAM | ! Matches BFL's 8 GB guidance: the int8 files in ComfyUI (published route match) BFL: the int8 ComfyUI files run on 8 GB GPUs such as the RTX 3070. Image model, int8 text encoder and VAE together are 11.0 GB; by our reading, on a card smaller than that ComfyUI keeps them on the GPU one at a time and needs system RAM for the rest. BFL gives no RAM figure. The prompt upsampler BFL says not to skip is a 9B model (6.72 GB w4a8, 9.75 GB int8 or 18.8 GB BF16 in ComfyUI); BFL does not say whether its 8 GB figure includes it. When it does not fit with the image model, run it on its own first (BFL's README) or switch it off in the ComfyUI template, with weaker results. |
| GeForce RTX 4060 Ti (16 GB) + 32 GB RAM | ! Matches BFL's 8 GB guidance: the int8 files in ComfyUI (published route match) BFL: the int8 ComfyUI files run on 8 GB GPUs such as the RTX 3070. Image model, int8 text encoder and VAE together are 11.0 GB; by our reading, on a card smaller than that ComfyUI keeps them on the GPU one at a time and needs system RAM for the rest. BFL gives no RAM figure. The prompt upsampler BFL says not to skip is a 9B model (6.72 GB w4a8, 9.75 GB int8 or 18.8 GB BF16 in ComfyUI); BFL does not say whether its 8 GB figure includes it. When it does not fit with the image model, run it on its own first (BFL's README) or switch it off in the ComfyUI template, with weaker results. |
| GeForce RTX 3090 / 3090 Ti (24 GB) + 64 GB RAM | ! Matches BFL's 8 GB guidance: the int8 files in ComfyUI (published route match) BFL: the int8 ComfyUI files run on 8 GB GPUs such as the RTX 3070. Image model, int8 text encoder and VAE together are 11.0 GB; by our reading, on a card smaller than that ComfyUI keeps them on the GPU one at a time and needs system RAM for the rest. BFL gives no RAM figure. The prompt upsampler BFL says not to skip is a 9B model (6.72 GB w4a8, 9.75 GB int8 or 18.8 GB BF16 in ComfyUI); BFL does not say whether its 8 GB figure includes it. When it does not fit with the image model, run it on its own first (BFL's README) or switch it off in the ComfyUI template, with weaker results. |
| GeForce RTX 5090 (32 GB) + 64 GB RAM | ! Matches BFL's 8 GB guidance: the int8 files in ComfyUI (published route match) BFL: the int8 ComfyUI files run on 8 GB GPUs such as the RTX 3070. Image model, int8 text encoder and VAE together are 11.0 GB; by our reading, on a card smaller than that ComfyUI keeps them on the GPU one at a time and needs system RAM for the rest. BFL gives no RAM figure. The prompt upsampler BFL says not to skip is a 9B model (6.72 GB w4a8, 9.75 GB int8 or 18.8 GB BF16 in ComfyUI); BFL does not say whether its 8 GB figure includes it. When it does not fit with the image model, run it on its own first (BFL's README) or switch it off in the ComfyUI template, with weaker results. NVIDIA measured about 4 seconds per 1 MP image on an RTX 5090 (distilled, int8, 16 steps, ComfyUI, upsampler not included). |
| Apple Silicon Mac (64 GB unified) | – Unknown No published FLUX 3 Klein route for this hardware. Black Forest Labs and NVIDIA publish no Apple silicon figures, and BFL's reference code installs for CUDA. A community MLX conversion for mflux exists; we have not verified it. The prompt upsampler BFL says not to skip is a 9B model (6.72 GB w4a8, 9.75 GB int8 or 18.8 GB BF16 in ComfyUI); BFL does not say whether its 8 GB figure includes it. When it does not fit with the image model, run it on its own first (BFL's README) or switch it off in the ComfyUI template, with weaker results. |
| Mac with M5 Max, 40-core GPU (128 GB unified) | – Unknown No published FLUX 3 Klein route for this hardware. Black Forest Labs and NVIDIA publish no Apple silicon figures, and BFL's reference code installs for CUDA. A community MLX conversion for mflux exists; we have not verified it. The prompt upsampler BFL says not to skip is a 9B model (6.72 GB w4a8, 9.75 GB int8 or 18.8 GB BF16 in ComfyUI); BFL does not say whether its 8 GB figure includes it. When it does not fit with the image model, run it on its own first (BFL's README) or switch it off in the ComfyUI template, with weaker results. |
| Mac Studio M3 Ultra (256 GB unified) | – Unknown No published FLUX 3 Klein route checked here for this device. |
| Mac Studio M3 Ultra (512 GB unified) | – Unknown No published FLUX 3 Klein route checked here for this device. |
Machine-readable: the same answers as JSON (CC BY 4.0, attribution "RigCheck").
Related hardware
Graphics card and Mac pages where this page's checker gives an answer, smallest memory first: graphics cards with 64 GB of system RAM, Macs with their unified memory, in the finder's quick mode (published route match; each page lists every model).
- RTX 3060 12 GB: Matches BFL's 8 GB guidance: the int8 files in ComfyUI
- RTX 4070 Ti: Matches BFL's 8 GB guidance: the int8 files in ComfyUI
- RTX 4060 Ti 16 GB: Matches BFL's 8 GB guidance: the int8 files in ComfyUI
- RTX 3090: Matches BFL's 8 GB guidance: the int8 files in ComfyUI
- RTX 4090: Matches BFL's 8 GB guidance: the int8 files in ComfyUI
- RTX 5090: Matches BFL's 8 GB guidance: the int8 files in ComfyUI
All graphics cards · All Macs · Any other machine
FAQ
How much VRAM does FLUX 3 Klein need?
Black Forest Labs says its int8 ComfyUI files run on 8 GB GPUs such as the RTX 3070. The image model is 11.1 GB in BF16, 5.83 GB in int8 and 5.28 GB in NVFP4; the text encoder adds 8.88 GB (BF16) or 4.86 GB (int8). BFL's own reference code peaked at 25–31 GB on a B200 at 1024×1024 without the prompt upsampler.
Can FLUX 3 Klein run on an 8 GB GPU?
Yes, by Black Forest Labs: use the int8 files in ComfyUI (it names the RTX 3070). The image model, text encoder and VAE together are 11.0 GB in int8, so expect ComfyUI to move components between the GPU and system RAM. BFL does not say whether the 8 GB figure includes the prompt upsampler (6.72–9.75 GB in ComfyUI): run it separately or switch it off, with weaker results. BFL gives no speed for an 8 GB card.
Which FLUX 3 Klein file should I use in ComfyUI?
For generation and editing, the step-distilled model (no base in the file name, 16 steps; BFL says about 6× faster than base, and its H200 example is 3 s vs 12 s). BFL points 8 GB cards and RTX 30 (no FP8) to int8; NVFP4 in its own code needs Blackwell, and its RTX benchmarks (RTX 5090 included) use int8; our suggestion: int8 up to about 16 GB, BF16 when the card has room for the 11.1 GB model plus the encoder. The base model is for fine-tuning and LoRA training.
Can I skip the prompt upsampler?
You can, but Black Forest Labs says results are noticeably weaker: the model is trained on the upsampler's captions. To save VRAM, use its int8 (9.75 GB) or w4a8 (6.72 GB) file, run it before the image model rather than alongside it, or use another LLM with BFL's system prompts.
Does FLUX 3 Klein run on a Mac?
Black Forest Labs and NVIDIA publish no Apple silicon figures, and BFL's reference code installs for CUDA. A community MLX conversion for mflux exists on Hugging Face (vanch007/FLUX.3-Image-Klein-mlx-8bit); we have not verified it.
Is FLUX 3 Klein free for commercial use?
Under the FLUX Kommunity License, yes for individuals and companies with less than $5M annual revenue, plus research and non-commercial use; larger organizations need BFL's self-hosted license or the API.
Sources
- Black Forest Labs: FLUX 3 Image Klein (parameters, 8 GB claim, NVIDIA-measured speeds, runtimes, license; read October 16, 2026)
- GitHub: black-forest-labs/flux-3-image-klein (models, GPU guidance, peak-memory table, upsampler, download sizes)
- Hugging Face file lists: flux-3-image-klein-comfy, flux-3-image-klein, -base and -pup (upsampler model card)
- ComfyUI and Diffusers modular_pipelines/flux3 source (checked October 16, 2026); community: Abiray and realrebelai GGUF, vanch007 MLX