Guides · both · run wan video locally

Open-weight Chinese video and image models, and the hardware the cards actually name

Published 16 Oct 2026. Updated 16 Oct 2026. Researched 16 Oct 2026.

Local video is the privacy argument that survives a terms-of-service change: the frames never leave the machine. It is also where blogs invent a tokens-per-second figure. This page sticks to the model cards fetched on 16 Oct 2026.

What the cards say

Wan2.2 TI2V-5B, from Wan-AI, is Apache 2.0. It does text-to-video and image-to-video at 720p and 24 fps. The card says a 5 second 720p clip can be under 9 minutes on one consumer GPU without extra optimization, and that the documented command, with offload, runs on a GPU with at least 24GB, for example an RTX 4090. It says 80GB lets you drop the offload flags. We did not time a run, so the 9 minutes stays their sentence, not our benchmark.

Qwen-Image is Apache 2.0 and 20B parameters in BF16. The card demonstrates Diffusers. It does not state a VRAM floor in the text we used, so this page does not invent one. The same family’s qwen-image and qwen-image-2512 rows are on the 7 Oct text-to-image board, licensed Apache 2.0 in that dataset. They are not the top of that board. Hosted Qwen-Image 3.0 Pro is a different, proprietary row and a QwenCloud price. Do not download the old weights and expect the hosted rank.

HunyuanVideo’s card sets the licence name to tencent-hunyuan-community and points at a licence file. The 22 Sep video board labels hunyuan-video-1.5 the same way. Community licences from Tencent are not “use it in a startup with no reading.” Open the file. We did not copy a VRAM number from that card.

Kolors is the Kuaishou image model. The card says academic use is open and commercial use requires a questionnaire to [email protected]. The code of the project is Apache 2.0. The weights are the part with the extra term. Kolors is absent from the overall image snapshot we stored.

The video board’s open-weight marker, on 22 Sep, is wan-v2.2-a14b with Apache 2.0. It is not near the top. A local model can still be the right tool when the constraint is data leaving the building, which the safety guide is about. The Qwen text local guide is the language-model version of this page. It is not a video install.

If 24GB is not on your desk, the hosted prices are the other option. Wan’s per-second rates and Seedream’s per-image rates are in the video and image pieces. BytePlus does not turn HunyuanVideo into a local install. It is a cloud bill.