Everything you load passes through system RAM first, and anything that does not fit in VRAM waits there. So RAM needs to hold the diffusion model, the text encoder, the VAE and Windows at the same time.
Rules of thumb
- 16 GB: SD 1.5, SDXL and anime checkpoints, SD 3.5 Medium, FLUX.2 [klein] 4B, Z-Image with an FP8 encoder. Tight for FLUX.1.
- 32 GB: FLUX.1 at 8-bit with an FP8 T5, Chroma, Krea 2, Wan 2.2 5B, HunyuanVideo 1.5. The comfortable minimum for most people.
- 48–64 GB: the big models at 8-bit: Qwen-Image, HiDream, Wan 2.1/2.2 14B (Wan 2.2 keeps two 14B models around), and FLUX.2 [dev] with its FP8 text encoder.
- 96 GB and up: MiniMax H3, 16-bit versions of the biggest models, or several big models loaded at once.
Every verdict page on this site gives a RAM figure for that exact combination: the file it recommends, plus the smallest 8-bit text encoder, plus about 6 GB for Windows and ComfyUI, rounded up to a normal RAM size.
Why offloading is slow
When a model spills out of VRAM, the GPU has to fetch weights from system RAM over PCIe during every step. The PCIe link, not the RAM, is usually the bottleneck — and some popular cards, such as the RTX 4060 Ti and RTX 5060 Ti, use only eight PCIe lanes, as do many laptops. ComfyUI's Dynamic VRAM (default on NVIDIA since March 2026) makes streaming much smarter, but more VRAM is still the real fix; more RAM just makes the slow path possible.
The page file
On older ComfyUI versions and non-NVIDIA setups, a full RAM means Windows falls back to the page file on your SSD: ComfyUI survives, but loading a model can take minutes. With Dynamic VRAM, ComfyUI maps model files from disk instead and relies on the page file much less.