Stable Diffusion 3.5 VRAM requirements (Medium and Large)
Short answer: Stability AI says SD 3.5 Medium needs 9.9 GB of VRAM, not counting the text encoders.[3] SD 3.5 Large's files total 27.6 GB, including a 16.3 GB transformer; Stability gives no VRAM figure for Large. Both use the same 11.3 GB of text encoders.[1][2]
Both models are released under the Stability AI Community License. Hugging Face lists 8,146,280,768 parameters for Large and 2,469,663,936 for Medium.[1][2]
Weights only: file sizes on Hugging Face
From the official file listings (1 GB = 1,000,000,000 bytes; split files added; checked November 4, 2026).[1][2]
| Component | SD 3.5 Large | SD 3.5 Medium |
|---|---|---|
| MMDiT transformer | 16.29 GB | 4.94 GB |
| T5-XXL text encoder (FP16) | 9.52 GB | 9.52 GB |
| CLIP-G text encoder | 1.39 GB | 1.39 GB |
| CLIP-L text encoder | 0.25 GB | 0.25 GB |
| VAE | 0.17 GB | 0.17 GB |
| Total (weights only) | 27.62 GB | 16.27 GB |
Both repos also ship an FP8 T5-XXL (t5xxl_fp8_e4m3fn.safetensors, 4.89 GB) for ComfyUI-style setups, about half the FP16 file.[1]
Memory while generating (official figure)
| Reference setup | Official VRAM |
|---|---|
| SD 3.5 Medium, Stability's hardware-compatibility comparison, "to unlock its full performance" | 9.9 GB, excluding text encoders[3] |
| SD 3.5 Large | Not published |
The text encoders can sit in system RAM or on the GPU, so add up to 11.3 GB (FP16) if you keep them on the GPU. We haven't measured Large ourselves, so we don't estimate it.
What this means for your GPU
- SD 3.5 Medium: 12 GB+ cards meet Stability's 9.9 GB figure with the text encoders offloaded. See what you can run on 12 GB.
- SD 3.5 Large: the transformer alone is 16.3 GB, so plan on a 24 GB card with the encoders offloaded, or rent one by the hour.
- Lighter option: SDXL VRAM (6.9 GB files).
Frequently asked questions
How much VRAM does Stable Diffusion 3.5 need?
Stability says Medium needs 9.9 GB, excluding text encoders. Large has no official figure; its transformer alone is 16.3 GB.
Can SD 3.5 Large run on 16 GB?
Not with its 16.3 GB transformer fully on the GPU at full precision. You need offload or a quantized build.
Every number on this page links to its source.
How we calculate + sources ⓘ
Sources
All sources checked on .
- Hugging Face, stabilityai/stable-diffusion-3.5-large (file listing, parameters, license; checked November 4, 2026). https://huggingface.co/stabilityai/stable-diffusion-3.5-large/tree/main
- Hugging Face, stabilityai/stable-diffusion-3.5-medium (file listing, parameters; checked November 4, 2026). https://huggingface.co/stabilityai/stable-diffusion-3.5-medium/tree/main
- Stability AI, "Introducing Stable Diffusion 3.5" (hardware compatibility; checked November 4, 2026). https://stability.ai/news/introducing-stable-diffusion-3-5