Can my PC run Stable Diffusion, Flux and Wan?
Wan 2.2 14B (A14B) has 14B parameters in each of its two experts and needs about 32 GB of graphics memory at full precision (FP16), with its working memory for a 5-second 480p video; about 18 GB at FP8; and about 13 GB as a GGUF Q4_K_M file, so it runs entirely on a 16 GB card or bigger.
| Model | Runs? | Time |
|---|
The model file and the working memory it needs on the card, with about 0.8 GB more kept for the driver and the screen. 8-bit (FP8 or Q8_0) looks almost the same as full precision; GGUF Q4 loses some fine detail.
| Precision | Model | On the card | Fits a card with |
|---|---|---|---|
| FP16 | 28.0 GB | 32.0 GB | 48 GB |
| FP8 | 14.0 GB | 18.0 GB | 20 GB |
| GGUF Q8_0 | 14.9 GB | 18.9 GB | 20 GB |
| GGUF Q6_K | 11.5 GB | 15.5 GB | 20 GB |
| GGUF Q5_K_M | 9.9 GB | 13.9 GB | 16 GB |
| GGUF Q4_K_M | 8.5 GB | 12.5 GB | 16 GB |
The text encoder, umT5-XXL, reads your prompt first, about 6.7 GB at 8 bits; ComfyUI moves it off the card before drawing, so it needs system memory rather than graphics memory.
Two 14B experts, one for the early steps and one for the late ones; ComfyUI loads one at a time, so the card needs room for one, and system memory for both. At 720p it takes about three times as long.
Time for a 5-second 480p video at 40 steps, at the largest precision that fits, in a PC with 32 GB of system memory.
| Graphics card | Runs? | Precision | Time |
|---|---|---|---|
| RTX 3060 12 GB | Partly in system memory | GGUF Q4_K_M | 55 min–1.8 hours |
| RTX 4060 (8 GB) | Partly in system memory | GGUF Q4_K_M | 42 min–83 min |
| RTX 4060 Ti 16 GB | On the card | GGUF Q5_K_M | 17 min–34 min |
| RTX 5060 Ti 16 GB | On the card | GGUF Q5_K_M | 16 min–31 min |
| RTX 4070 (12 GB) | Partly in system memory | GGUF Q4_K_M | 25 min–51 min |
| RTX 5070 (12 GB) | Partly in system memory | GGUF Q4_K_M | 23 min–45 min |
| RTX 4070 Ti SUPER (16 GB) | On the card | GGUF Q5_K_M | 8 min–17 min |
| RTX 5070 Ti (16 GB) | On the card | GGUF Q5_K_M | 8 min–17 min |
| RTX 3090 (24 GB) | On the card | FP8 | 10 min–20 min |
| RTX 4080 SUPER (16 GB) | On the card | GGUF Q5_K_M | 7 min–14 min |
| RTX 5080 (16 GB) | On the card | GGUF Q5_K_M | 7 min–14 min |
| RTX 4090 (24 GB) | On the card | FP8 | 4 min–8 min |
| RTX 5090 (32 GB) | On the card | FP8 | 4 min–7 min |
| Radeon RX 9070 XT (16 GB) | On the card | GGUF Q5_K_M | — |
| Radeon RX 7900 XTX (24 GB) | On the card | FP8 | — |
| Arc B580 (12 GB) | Partly in system memory | GGUF Q4_K_M | — |
AMD and Intel cards: whether it fits, but no time, which depends too much on the software.
| Made by | Wan (Alibaba) |
|---|---|
| Released | July 2025 |
| Makes | 5-second 480p videos |
| Parameters | 14B per expert, two experts |
| Licence | Apache 2.0 |
Its files are on Hugging Face. ComfyUI runs it with a ready-made workflow; GGUF files need the ComfyUI-GGUF add-on. Bottleneck is not connected with Wan (Alibaba).
Image and video models are limited by the graphics card’s arithmetic rather than its memory speed, so each card’s speed here comes from its measured arithmetic in llama.cpp’s benchmark (the same figure the AI model calculator uses for reading prompts), and each model’s time per step is anchored to measured times: FLUX.1 dev at about 1.85 steps a second on an RTX 4090, SDXL at about 4.2 a second on an RTX 3090, and Wan 2.2 5B taking about nine minutes for a 720p clip on an RTX 4090, its makers’ own figure. Other models scale from those by size. Memory is the model at each precision plus its working memory at the usual resolution, with about 0.8 GB kept for the driver and the screen. Speeds are for NVIDIA cards; AMD and Intel cards run these models through ROCm, ZLUDA, DirectML or Intel’s own builds, whose speed varies too much with the software to estimate. A real time is usually within the range shown; speed-up add-ons such as fewer-step LoRAs, TensorRT or SageAttention make it faster.
Bottleneck’s free app reads your graphics card, how much of its memory other programs are holding, your memory’s speed and channels, and the settings that slow a PC down, and says how to fix each. Its AI models screen lists the chat models your PC runs, and checks one you already run with Ollama or LM Studio.
The free version gives each check’s verdict, with no time limit and no account. Pro, $9 once, explains the cause and the fix. Compare the versions