Wan 2.2 14B is two 14B models: a high-noise expert for the first half of the steps and a low-noise expert for the second half. In FP8 each one is 14.3 GB, too big to sit in 16 GB together with the working memory. The usual advice is to take a GGUF that fits. I timed both on my RTX 5060 Ti 16 GB.
The result
Minutes per clip: text to video, 832×480, 81 frames (about 5 seconds), 20 steps (10 + 10). Shorter is better.
| File type | Size | Fits in 16 GB? | Per clip | Peak VRAM | Peak RAM |
|---|---|---|---|---|---|
| FP8 | 14.3 GB ×2 | no | 19.1 min | 13.3 GB | 28.6 GB |
| GGUF Q5_K_M | 10.8 GB ×2 | yes | 25.7 min | 14.0 GB | 33.1 GB |
| GGUF Q8_0 | 15.4 GB ×2 | no | 25.3 min | 13.1 GB | 32.9 GB |
Timed runs after 1 warm-up run. The warm-up runs were within 1% of the timed runs. Peak VRAM and peak RAM are the memory in use on the whole card and the whole PC during the run. The PC has 32 GB of RAM.
What it means
- FP8 was the fastest, even though it does not fit. ComfyUI's Dynamic VRAM (on by default on NVIDIA since March 2026) streams the part that does not fit from system RAM, and RTX 40/50 cards compute FP8 in hardware. GGUF weights have to be unpacked at every step.
- The GGUF that fits was not faster. Q5_K_M fits in VRAM but took 25.7 min. Q8_0, which does not fit either, took about the same.
- 32 GB of RAM is the tight spot. With the GGUF files the RAM was completely full during every run (33.1 GB in use); with FP8 it peaked at 28.6 GB. The warm-up runs took about as long as the timed ones, so the experts seem to be reloaded on every run. With 64 GB of RAM it will likely be faster; I have not measured that yet.
- My advice for 16 GB on RTX 40/50: use the FP8 files from Comfy-Org with an up-to-date ComfyUI. Use GGUF only if you are on an older GPU, AMD or Intel, or an old ComfyUI version.
If 20 minutes per clip is too slow for you, a rented 48 GB or bigger card runs both experts in VRAM with room to spare. Rent one by the hour on Vast.ai (affiliate link: I get a small share, at no extra cost to you).
This is the opposite of the rule most guides still repeat ("take the biggest GGUF that fits"). On new NVIDIA cards and a current ComfyUI, it no longer holds for this model. See also FP8 vs GGUF with FLUX.1 dev and how much system RAM you need.
My setup
- GPU: NVIDIA GeForce RTX 5060 Ti 16 GB, desktop. System RAM: 32 GB. Windows, models on an NVMe SSD.
- ComfyUI 0.25.0 portable, PyTorch 2.12 with CUDA 13.0, default launch options.
- Files: Comfy-Org
wan2.2_t2v_high/low_noise_14B_fp8_scaled; QuantStackWan2.2-T2V-A14B-HighNoise/LowNoiseGGUF (Q5_K_M, Q8_0) loaded with ComfyUI-GGUF. - Text encoder UMT5-XXL FP8, VAE Wan 2.1. Sampler euler, scheduler simple, CFG 3.5, shift 8, two KSamplerAdvanced nodes (steps 0–10 and 10–20). No speed-up LoRA.
- Time is measured by ComfyUI from start to end of the job: loading both experts, sampling and VAE decode.
- Date: 2026-09-27.
More results from my card: RTX 5060 Ti 16 GB and all measured and reported results. Ran the same test on another card? Send me your numbers.