Run AI Image Models Locally: What Fits Your GPU
How much video memory (VRAM) your graphics card has decides which AI image models you can run at home. By the makers’ own figures, FLUX 3 Image Klein has an INT8 build that runs on an RTX card with as little as 8 GB of VRAM, FLUX.2 [klein] 4B fits in about 13 GB, Stable Diffusion 3.5 Medium needs 9.9 GB before its text encoders, and Z-Image-Turbo fits within 16 GB. This page puts those numbers side by side and shows how to read them.
What the makers say each model needs
Every figure in this table is the maker’s own statement from its announcement or model card, not an independent measurement. The numbers are not measured the same way, so treat them as a starting point rather than a promise.
| Model | Size | Memory the maker states | License as the maker describes it |
|---|---|---|---|
| FLUX 3 Image Klein (Oct 15, 2026) | 5.6B parameters | Generator weights: 11.2 GB in BF16, 5.6 GB in FP8. The INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM, through ComfyUI or Diffusers. | FLUX Kommunity License: free commercial use for individuals and companies with annual revenue under $5M, plus research and non-commercial use. |
| FLUX.2 [klein] 4B | 4B parameters | Fits in about 13 GB of VRAM; the card names the RTX 3090 and 4070 and above. | Apache 2.0. |
| Stable Diffusion 3.5 Medium | 2.5B parameters | 9.9 GB of VRAM, excluding text encoders. | Stability AI Community License: free for non-commercial use, and free for commercial use under $1M in annual revenue. |
| Z-Image-Turbo | 6B parameters | Fits comfortably within 16 GB of VRAM on consumer devices; the card says 8 NFEs (function evaluations), and its example sets 9 steps. | Apache 2.0, per the license field on the model card. |

How to read these numbers
- Weights are not the whole bill. Black Forest Labs says its parameter counts and weight sizes are for the image generator only, and Stability says its 9.9 GB figure leaves out the text encoders. The text encoder and other parts need memory too, so plan for more than the headline number.
- Precision changes the size. The same model can ship as BF16, FP8 or INT8 builds. Black Forest Labs lists 11.2 GB for BF16 and 5.6 GB for FP8 weights of FLUX 3 Image Klein, and says the INT8 build is the one meant for 8 GB cards.
- Offloading trades speed for memory. The FLUX.2 [klein] 4B model card’s own example calls enable_model_cpu_offload() to save VRAM. The card does not say how much time that costs.
- Step counts change speed. The ComfyUI guide for FLUX 3 Image Klein says its distilled checkpoint runs in 4 to 16 steps and the base checkpoint defaults to 50, and the Z-Image-Turbo card says 8 NFEs.
A quick way to choose
- 8 GB card: start with the FLUX 3 Image Klein INT8 build, the one Black Forest Labs says targets 8 GB RTX cards.
- Around 10 GB: Stable Diffusion 3.5 Medium is the smallest of the four at 2.5B parameters, but Stability’s 9.9 GB figure leaves out the text encoders.
- 13 GB: FLUX.2 [klein] 4B, with the permissive Apache 2.0 license.
- 16 GB: Z-Image-Turbo is the one its makers say fits comfortably, and its card says it needs only 8 NFEs.
These are the makers’ claims about what fits. Actual memory use depends on image size, the text encoder you load and the software you use, so check your own numbers before you download large files.
Licenses matter as much as memory
Two of these four models use licenses with conditions. FLUX 3 Image Klein’s license, as Black Forest Labs describes it, allows free commercial use below $5M in annual revenue, and Stable Diffusion 3.5’s community license does so below $1M. FLUX.2 [klein] 4B and Z-Image-Turbo are Apache 2.0. Read the full license text for any model before you use its images commercially. This page summarizes what the makers say and is not legal advice.

How to run them
ComfyUI is the most direct route for the FLUX 3 Image Klein weights. The ComfyUI guide to FLUX 3 Image Klein says its templates need ComfyUI 0.41.0 or later and lists text-to-image and image-edit workflows, with distilled versions of each. The FLUX.2 [klein] 4B and Z-Image-Turbo model cards also include Diffusers code examples.

Where these numbers come from
Every figure above is quoted from the maker’s own announcement or model card, linked in the table, and was checked on November 18, 2026. Model sizes, builds and licenses change quickly, so check the linked page before you download a large file. For more guides like this one, browse the Run It Locally category.
Header image: Black Forest Labs via bfl.ai/blog/flux-3-image-klein.