Wan 2.2 Hits Local Rigs: 720p Video on 24GB of VRAM

Open-source video generation just landed on hardware you can actually buy. Wan 2.2 runs locally, and its lightweight TI2V-5B model produces 720p clips at 24fps on a single graphics card, with the heavier 14B model delivering markedly better results for those with the VRAM to feed it. For anyone who has watched local AI image generation mature and wondered when video would follow, this is the moment it became practical on a home rig.

Quick Answer

Wan 2.2's TI2V-5B model generates 720p video at 24fps on a single RTX 4090's 24GB of VRAM, producing a roughly 5-second clip in about 9 minutes. The model is light enough to run on cards with around 8GB at reduced settings, while the larger 14B model needs 24GB or more and rewards it with stronger output.

What Wan 2.2 brings to local rigs

Wan 2.2 ships in two practical flavours. The TI2V-5B is the headline for home users: a 5-billion-parameter hybrid model that handles both text-to-video and image-to-video at 720p and 24fps, and it is one of the fastest models in that resolution-and-framerate bracket you can run yourself. The larger A14B uses a mixture-of-experts design, splitting the work between a high-noise expert for overall layout and a low-noise expert for detail, which is why its output looks noticeably more refined.

The practical headline is that the 5B model runs comfortably on a single RTX 4090, generating a 5-second 720p clip in roughly 9 minutes. That is slow by gaming standards but transformative for local video, because until recently this kind of generation meant renting cloud compute by the minute.

What you can run on your VRAM

Your graphics card decides which door is open. The 5B model is the accessible one, needing around 8GB of VRAM to run at reduced resolution and quality, which brings local video into reach for far more people. Step up to a 24GB card like the RTX 4090 and you can run the 5B model at full 720p smoothly, and you gain the headroom to load the 14B model, which needs 24GB or more for serious work.

The gap between the two models is real. The 5B is fast and capable, the 14B produces some of the most impressive output from any open video model, and the cost of that quality is VRAM and time. If video generation is the reason you are building or upgrading, VRAM is the spec to prioritise above almost everything else, because it sets the hard ceiling on what you can load at all.

A capable, well-cooled rig matters here too, since these generations run the card flat-out for minutes at a time. Purpose-built rigs that handle sustained GPU load without throttling are listed in the AI-ready PC range at Evetech, and to pick the card itself, the GPU best sellers at Evetech reflect which high-VRAM options local buyers are choosing.

Frequently Asked Questions

What VRAM do I need to run Wan 2.2?

The TI2V-5B model runs on roughly 8GB of VRAM at reduced settings and comfortably at full 720p on a 24GB card. The larger 14B model needs 24GB or more, so a card like the RTX 4090 is the practical entry point for the higher-quality model.

How long does a clip take to generate?

On a single RTX 4090, the 5B model produces a 5-second 720p clip in about 9 minutes. Times scale with resolution, clip length, and which model you run, so the 14B model takes longer for its higher quality.

What is the difference between the 5B and 14B models?

The 5B is faster and runs on modest hardware, while the 14B uses a mixture-of-experts design for noticeably more refined output. The 14B needs far more VRAM, so the choice comes down to whether you have the card to feed it.

Can Wan 2.2 do image-to-video?

Yes. The TI2V-5B model supports both text-to-video and image-to-video, so you can animate a still image or generate a clip from a prompt. This flexibility is part of why the 5B model is so popular for local use.

Is an 8GB card actually usable for this?

It works for the 5B model at reduced resolution and quality, which is enough to experiment and learn the pipeline. For full 720p output and access to the 14B model, a 24GB card is the sensible target.