It depends less on the GPU and more on which model you want. A 4B model like FLUX.2 [klein] 4B is happy on 8 GB. FLUX.2 [dev] and MiniMax H3 do not fit in full quality even on a 32 GB RTX 5090. Here is what each common amount of VRAM gets you, calculated from the actual file sizes on Hugging Face.
8 GB
Runs well (8-bit or better): FLUX.2 klein 4B, Mage-Flow, Lumina 2.0, SD 3.5 Medium, SDXL, Illustrious / Pony, SD 1.5, Wan 2.1 1.3B.
Runs with a compressed GGUF: Qwen-Image 2.1, Z-Image Turbo, Z-Image, ERNIE-Image, SD 3.5 Large, Chroma1-HD, Wan 2.2 5B. Full list for 8 GB →
12 GB
Runs well (8-bit or better): FLUX.2 klein 9B, FLUX.2 klein 4B, Qwen-Image 2.1, Z-Image Turbo, Z-Image, ERNIE-Image, HiDream-O1, Mage-Flow, Lumina 2.0, SD 3.5 Large, SD 3.5 Medium, Chroma1-HD, SDXL, Illustrious / Pony, SD 1.5, Wan 2.1 1.3B, Wan 2.2 5B.
Runs with a compressed GGUF: FLUX.1 dev, FLUX.1 schnell, FLUX.1 Kontext, FLUX.1 Krea, FLUX.1 Fill, Krea 2, Boogu-Image, HunyuanVideo 1.5. Full list for 12 GB →
16 GB
Runs well (8-bit or better): FLUX.1 dev, FLUX.1 schnell, FLUX.1 Kontext, FLUX.1 Krea, FLUX.1 Fill, FLUX.2 klein 9B, FLUX.2 klein 4B, Krea 2, Qwen-Image 2.1, Z-Image Turbo, Z-Image, Boogu-Image, ERNIE-Image, HiDream-O1, Mage-Flow, Lumina 2.0, SD 3.5 Large, SD 3.5 Medium, Chroma1-HD, SDXL, Illustrious / Pony, SD 1.5, Wan 2.1 1.3B, Wan 2.2 5B, HunyuanVideo 1.5.
Runs with a compressed GGUF: Qwen-Image, Ideogram 4, HiDream-I1 Full, HiDream-I1, HunyuanImage 2.1, Wan 2.1 14B, Wan 2.1 I2V 480P, Wan 2.2 T2V, Wan 2.2 I2V, HunyuanVideo 13B, LTX-Video 13B. Full list for 16 GB →
24 GB
Runs well (8-bit or better): FLUX.1 dev, FLUX.1 schnell, FLUX.1 Kontext, FLUX.1 Krea, FLUX.1 Fill, FLUX.2 klein 9B, FLUX.2 klein 4B, Krea 2, Qwen-Image, Qwen-Image-Edit, Qwen-Image 2.1, Z-Image Turbo, Z-Image, Ideogram 4, Boogu-Image, ERNIE-Image, HiDream-O1, Mage-Flow, Lumina 2.0, HiDream-I1 Full, HiDream-I1, SD 3.5 Large, SD 3.5 Medium, Chroma1-HD, SDXL, Illustrious / Pony, SD 1.5, HunyuanImage 2.1, Wan 2.1 14B, Wan 2.1 1.3B, Wan 2.1 I2V 480P, Wan 2.1 I2V 720P, Wan VACE 14B, Wan 2.2 T2V, Wan 2.2 I2V, Wan 2.2 5B, Wan 2.2 Animate, Wan Animate 2, Wan 2.2 S2V, SCAIL-2, HunyuanVideo 13B, LTX-Video 13B, HunyuanVideo 1.5.
Runs with a compressed GGUF: FLUX.2 dev, LTX-2, LTX-2.3, LTX-2.5, MiniMax H3 Pruned. Full list for 24 GB →
32 GB
Runs well (8-bit or better): FLUX.1 dev, FLUX.1 schnell, FLUX.1 Kontext, FLUX.1 Krea, FLUX.1 Fill, FLUX.2 klein 9B, FLUX.2 klein 4B, Krea 2, Qwen-Image, Qwen-Image-Edit, Qwen-Image 2.1, Z-Image Turbo, Z-Image, Ideogram 4, Boogu-Image, ERNIE-Image, HiDream-O1, Mage-Flow, Lumina 2.0, HiDream-I1 Full, HiDream-I1, SD 3.5 Large, SD 3.5 Medium, Chroma1-HD, SDXL, Illustrious / Pony, SD 1.5, HunyuanImage 2.1, Wan 2.1 14B, Wan 2.1 1.3B, Wan 2.1 I2V 480P, Wan 2.1 I2V 720P, Wan VACE 14B, Wan 2.2 T2V, Wan 2.2 I2V, Wan 2.2 5B, Wan 2.2 Animate, Wan Animate 2, Wan 2.2 S2V, SCAIL-2, HunyuanVideo 13B, LTX-Video 13B, HunyuanVideo 1.5, LTX-2, LTX-2.3, LTX-2.5, MiniMax H3 Pruned.
Runs with a compressed GGUF: FLUX.2 dev, MiniMax H3. Full list for 32 GB →
If you are buying a GPU for this
- 8 GB works for SDXL and its anime cousins, SD 1.5, the small FLUX.2 [klein], Z-Image and the 5B Wan. Big models need heavy compression.
- 12 GB is where FLUX.1 becomes comfortable with a Q5 file, and Qwen-Image 2.1 runs at 8-bit.
- 16 GB is the sweet spot today: FLUX.1 and Krea 2 at 8-bit, Wan 14B video with a Q5 file.
- 24–32 GB runs almost every image model at 8-bit, and video at higher resolutions.
Memory matters more than speed here. A slower GPU with more VRAM holds entirely what a faster one has to stream or compress. Compare GPUs model by model →
How the verdicts are worked out: how the numbers work.