Best open source AI video generator: Wan 2.2, the GPU it needs, and the open model the US can't use

The best open source AI video generator is Wan 2.2, because it's the open video model a US user can actually use, under Apache 2.0, and it runs on consumer GPUs: the 5B version from 8 GB with ComfyUI's offloading, the 14B at 640x640 on a 24 GB card. The other big open release, MiniMax H3, has a licence that excludes the United States, and Alibaba's newer Wan 3.0 has no open weights at all. Licences, VRAM figures and prices below come from the makers' pages, ComfyUI's and Runpod's docs, and our own GPU pages, read November 1 to 8, 2026.

The 24 GB tier is where most people should start: it covers the 14B models at a modest resolution and the 5B model with room to spare. Runpod's guide says its boundaries are derived from published figures rather than measured on each card, so treat them as starting points. For every hosted model compared job by job, see the best AI video generator guide.

Which open video models can you legally use in the US?

Open weights and an open licence aren't the same thing, and for video in late 2026 the licence decides it.

• Wan 2.2 is under Apache 2.0, a permissive licence that allows commercial use. Alibaba's Wan2.2 README lists the variants: T2V-A14B and I2V-A14B at 480P and 720P, TI2V-5B (text and image to video at 720P and 24 fps, 1280x704 or 704x1280, so vertical works), S2V-14B (speech to video) and Animate-14B.
• MiniMax H3 launched on July 31, 2026 as an open model with native stereo sound, up to 15 seconds at 2K, per MiniMax. But its licence file defines "Excluded Territories" as "the European Union, the United Kingdom, the Republic of Korea and the United States of America". Renting a GPU in another country doesn't change where you are. MiniMax says people in those places can contact it about a separate licence.
• Wan 3.0 has no weights: Alibaba's Wan-AI page on Hugging Face lists Wan 2.1, 2.2 and Wan-Dancer, and the Wan-Video GitHub has no Wan 3.0 repo. You can only use it on Alibaba's app and API or on a platform such as Higgsfield (is Wan 3 free).

What GPU does Wan 2.2 need?

It depends on which Wan 2.2 and at what resolution, and the three sources you'll read measure different setups, so their numbers look contradictory until you see which setup each one means.

The 5B model is the one to learn on. ComfyUI's docs say it "should fit well on 8GB vram with the ComfyUI native offloading". Alibaba's own README asks for at least 24 GB for its plain sample command, for example an RTX 4090, and says that command makes a 5-second 720P clip in under nine minutes on one consumer GPU. Both are right: ComfyUI moves parts of the model to system memory, and Alibaba's script doesn't.

The 14B models are what people mean by Wan 2.2 quality. Runpod's guide says they need 24 GB at 640x640 with fp8 weights; fp16 weights need a 48 GB card, and 720p without offloading belongs on an 80 GB card. Alibaba's README says its 14B sample commands need at least 80 GB. Wan 2.2 14B is two experts rather than one network, so the guide also says to rent a machine with enough system RAM for the idle one.

Our cheapest cloud GPU for ComfyUI and Wan 2.2 page matches each tier to the cheapest card that fits, and there's no Wan entry in our VRAM calculator yet, so use these published tiers rather than a calculator figure.

What does one clip cost on a rented GPU vs an app?

GPU time per clip is tiny; the bill around it isn't. Runpod's guide timed a 5-second Wan 2.2 14B clip at 640x640 on a 24 GB card at about 71 seconds with the 4-step Lightning LoRA and about 513 seconds at 20 steps, after the first run. Here's how that compared with paying per clip, from our app, API or rented GPU cost breakdown and Higgsfield's prices:

The rented-GPU rows assume each clip is rendered twice, half an hour of setup per session and 100 GB of models on a network volume, the inputs in our cost breakdown; change them and the totals move. Runpod also lists a different Wan 2.2 endpoint, "with LoRAs", at $0.35 per 5 seconds. GPU prices move every hour, so don't plan on these: today's rates for every card are in our live cloud GPU price table.

The pattern holds whatever the hourly rate: at about 20 clips a month, renting costs as much as an app or more; at 100, it costs a fraction. Storage and setup are fixed costs, so small batches carry them badly.

When an app beats running it yourself

Open source isn't the answer for everyone, and it's worth being honest about the trade:

• Sound. Alibaba's README doesn't mention generated audio for Wan 2.2's text and image models; S2V-14B animates from a speech track you supply. Hosted Wan 3.0, Kling and Seedance make sound with the picture.
• Resolution and length. Wan 2.2 tops out at 720P, and the README's example is a 5-second clip. Hosted Wan 3.0 goes to 1080p and 30 seconds per Alibaba.
• Your time. The first session on a rented GPU goes to downloading tens of gigabytes of models and setting up a workflow. An app is a prompt box.

So: rent a GPU for Wan 2.2 if you make 50 or more clips a month, want full control of the workflow (LoRAs, your own fine-tunes) or need a licence you can read in full. Use a hosted model for a handful of clips, for sound, or for 1080p.

Where does Higgsfield fit?

Higgsfield is the no-GPU side of this choice. Its catalogue (Nov 8) hosts Alibaba's Wan 2.6, 2.7, 3.0 and 3.0 Prime, not Wan 2.2, alongside Kling, Seedance, Veo, MiniMax H3 and Grok Imagine. Wan 3.0 cost 9 credits for 5 seconds at 720p on November 3, about $0.30 on Ultra billed yearly ($99 a month for 3,000 credits) and about $0.44 on Plus monthly ($59 for 1,200), our arithmetic. Higgsfield also runs MiniMax H3 as a hosted service, which is a different thing from running its open weights. Prices change often, so check the price shown to you before paying. For the hosted models priced side by side, see Kling AI alternatives and the Kling hub.

The best AI video generator for other jobs

Each page below answers one job from the same checked facts. All of them, grouped by job, are in our best AI video generator guide.

• Best free AI video generator: Google Flow's 50 daily credits are the biggest free allowance; Dreamina's 120 a day is the US runner-up.
• Best image to video AI: Seedance 2.5 to put your photo's subject in a new scene; Kling 3.0 when the photo is already the first frame.
• Best text to video AI: Kling 4.0: up to 30 seconds per prompt, multi-shot and stereo sound; Google Flow and Dreamina for free tries.
• Best realistic AI video generator: Seedance 2.5 for realistic people, Kling 3.0 for native 4K detail, Veo 3.1 for realistic sound.
• Best AI video generator for long videos: Kling 4.0 makes 30-second takes, the longest, at a third of Seedance 2.5's credits; Seedance can extend.
• Best AI animation generator: Kling 3.0 animates your drawing from its first frame; Kling 3.0 Omni and 4.0 keep one character across scenes.

What we couldn't verify

• Wan 2.2 render times on any specific card we rented. We haven't timed Wan 2.2 ourselves; the 71-second and 513-second figures are Runpod's, and the under-nine-minutes figure is Alibaba's.
• MiniMax H3's VRAM needs. MiniMax publishes no VRAM figure, and for US readers the licence makes it moot.
• Wan 2.6 weights. Higgsfield labels Wan 2.6 open-weight, but we found no Wan 2.6 weights on Alibaba's Hugging Face or GitHub pages.
• Today's GPU prices. The dollar figures above use prices read Nov 1 to 3; the live table has today's.
• Video quality of Wan 2.2 against hosted models. We haven't run Wan 2.2, so we make no claim about how its clips compare.

If you're unsure whether you'll make enough clips to justify a GPU, price a month of clips on a hosted model first; if the number keeps growing, that's the signal to rent.