Wan 2.2 GGUF

GGUF is a single-file format for quantised model weights — the file you download when you want to run a model on your own machine. A Wan 2.2 GGUF is therefore not something this site serves: wan2-6.net is a hosted workspace and does not distribute weights or offer downloads. This page exists because the question is common enough to deserve a straight answer about which route fits you.

What the quantisation actually buys you

Full-precision weights are large and were designed for datacentre hardware. A GGUF file stores the same model at reduced numerical precision, which shrinks the file and, more importantly, shrinks the memory it needs while it runs. Quantisation levels trade size against fidelity: a smaller quantisation runs on a weaker card, and it will show that in fine detail and motion stability.

Practically, the constraint that decides everything is VRAM, not disk. Fitting the weights plus the attention workspace plus the VAE inside your card's memory is the whole game, and it is why the same model arrives as several quantisation levels rather than one.

When local is genuinely the right call

• You are running many generations and the marginal cost matters. After the hardware is paid for, local generation has no per-clip cost.
• You need to keep the material private. Nothing leaves the machine.
• You want to modify the pipeline — custom nodes, your own LoRA, a workflow you control end to end.

When a hosted workspace wins

• You do not have the card. A quantised model still wants a serious GPU; a laptop that runs the file may take many minutes per clip.
• You want to iterate, not maintain. Drivers, CUDA versions, node graphs and dependency conflicts are the tax on local installs, and they are paid before every first render.
• You need to know the cost before you commit. On the hosted route the credit cost is shown beside the form before you submit, so a test clip is a decision rather than a guess. Credit prices are on pricing.

Choosing without wasting a weekend

The cheap experiment is to start hosted and see whether the prompts you care about produce anything worth keeping, then decide whether the volume justifies hardware. Doing it the other way round means spending a weekend on an install before you know whether the model suits your idea.

If you decide to go hosted, the routes here are text to video, image to video, and the Wan 2.2 model page for what this generation exposes. If your interest is specifically quantised image-to-video, see Wan 2.2 I2V GGUF.

Questions people ask

Do you provide the GGUF file? No. This site runs models for you; it does not distribute weights.

Which quantisation should I pick? The largest one that fits your card's VRAM, leaving headroom for the attention workspace. If you must choose smaller, expect softer detail.

How much VRAM do I need? It depends on the quantisation, the resolution and the clip length. Short, low-resolution tests fit in far less than the sizes people quote online for full clips.

Is a hosted run the same model? The live form names the routes and models actually available; check it rather than assuming a match with a specific community weight file.