Stated GPU and VRAM Requirements
Running FLUX 3 Klein (full name FLUX 3 Image Klein) locally requires a capable graphics card. Only two hardware specifications are published in vendor documentation as of October 16, 2026.
| Source | Configuration | Stated Minimum Hardware | Documentation Link |
|---|---|---|---|
| Black Forest Labs (BFL) | Model documentation | NVIDIA RTX 3090 | BFL Documentation |
| ComfyUI | INT8 version | GPU with 8 GB of VRAM | ComfyUI Tutorial |
Black Forest Labs, the German company that makes the FLUX model family, gives an NVIDIA RTX 3090 as the minimum GPU for the model, whose weights are in the Hugging Face repository. ComfyUI’s documentation describes an INT8 version and says it runs on GPUs with 8 GB of VRAM. You can read more about setting up this workflow on our ComfyUI guide or check our weights download page for access instructions.
Understanding the Two-Part Architecture
Memory usage in local image generation depends heavily on the size of the model components that must be loaded into your hardware. FLUX 3 Klein is not a single file. It consists of a 5.6 billion parameter image model paired with a separate Qwen3-VL-4B text encoder.
In general terms, a text encoder turns the prompt into a form the image model works from, and both parts take memory while they are loaded. The parameter counts alone do not tell you the VRAM a setup needs: precision (for example INT8), offloading and the output size all change it. In addition, the model card notes that FLUX 3 Klein is meant to be used with a prompt upsampler (prompt expander), because short prompts used directly give mediocre results. If you run a local prompt expander alongside the main components, that tool may also require hardware resources.
Why These Numbers Are Vendor Claims
The RTX 3090 baseline and the 8 GB VRAM figure for INT8 are direct vendor claims from BFL and ComfyUI. We have not independently benchmarked local memory consumption across different operating systems, batch sizes, or output resolutions.
Actual memory usage during local inference can vary based on how your software manages the 5.6B model and the Qwen3-VL-4B encoder, as well as the resolution you select (up to 4K). Furthermore, we found no statement from Black Forest Labs on Apple Silicon or Mac support as of October 16, 2026. You can read more on our Mac compatibility page.
What to Do If You Have Less VRAM
If your computer has less than 8 GB of VRAM, or if you do not have a compatible GPU at all, you do not need to upgrade your hardware to use the model. You can run generations in the cloud instead of downloading the gated weights.
- Use Klein Studio: This website (flux3klein.org) is an unofficial service that uses Black Forest Labs’ API. It lets you generate images from text and edit one uploaded image directly in your browser. Every visitor gets 3 free images a day at 1K resolution with no sign-up required. You can start generating on the home page.
- Use the API directly: Developers and technical users can send requests straight to BFL’s servers, bypassing local hardware limits entirely. See our API overview for details on endpoint structure and third-party availability, or review our license guide if you plan to use the outputs for business purposes.