There is no free way to run Wan 2.5, and the honest version of this guide has to start there. Alibaba published weights for Wan 2.1 and Wan 2.2, then stopped: Wan 2.5 shipped as a hosted API in September 2025 with no public weights, and Wan 3.0 followed the same pattern. We checked the Wan-AI organization on Hugging Face on 2026-08-28 and the newest published repositories are still in the 2.2 line. So if "wan 2.5 free" means self-hosting, the answer is that the model you can actually download is Wan 2.2. This guide covers what that gets you, what it costs in hardware, and the hosted math for 2.5 if you need the newer model.
TL;DR
- Wan 2.5 has no published weights, so it cannot be self-hosted. Wan 2.2 (Apache 2.0) is the free, self-hostable one.
- The real price is hardware: a 24GB-VRAM-class GPU is the practical minimum for the full-quality variants, plus real setup time.
- Every other "free" option in AI video is a watermarked ration: Kling's free tier, Hailuo's, Pika's all stamp their output.
- No GPU? Hosted Wan 2.5 on 8frame is 102 credits per clip, about $1.02 at pack rate, roughly 9 clips on the $19 Starter plan.
Why Wan is the only genuinely free route
Free AI video usually means a demo. Kling's free tier is roughly three watermarked generations a day. Hailuo and Pika watermark their free output too. Those tiers exist to let you evaluate a closed model before paying, and they're fine at that job, but none of them produce a clip you can ship.
Wan 2.2 is different in kind, not in degree. Open weights means the actual model files are published for anyone to download and run, the way open weights vs closed models lays out. There's no watermark because there's no company between you and the model. No daily cap because the compute is yours. If "free" in your search meant "free forever, and I own the output," self-hosted Wan 2.2 or LTX-2.5 are the honest answers in 2026, and what is Wan AI covers how Alibaba's family got here.
So the question isn't whether Wan 2.5 is free. It's whether you own the machine it takes.
What self-hosting actually takes
Here's the entry ticket, stated plainly.
The GPU. The Wan family's small variants squeeze into 4 to 6 GB of VRAM and the mid-size ones target 8 to 12 GB, but quality lives in the large variants, and those want up to 22 to 26 GB for full-precision runs. In practice a 24GB-class consumer card is the floor for output you'd compare to the hosted version. Aggressive quantization and CPU offloading can shrink that, at a cost in speed and fidelity you'll notice.
The setup. Model weights measured in tens of gigabytes, a ComfyUI-style node pipeline or the reference inference scripts, and a working grasp of quantization and offloading to fit big checkpoints onto smaller cards. Budget an afternoon if everything goes right and a weekend if it doesn't.
The wait. On consumer hardware a single short clip takes minutes, not seconds. That's fine for a nightly render queue. It's painful for iteration, where you want ten variants before lunch.
None of this is a knock on Wan. It's the normal price of owning a pipeline, and if you have the card and enjoy the tinkering, self-hosted Wan 2.2 is a legitimately great deal: unlimited free b-roll for the cost of electricity.
The self-host route in outline
If you're going local, the path looks like this. Deliberately an outline, not a walkthrough, because the tooling moves monthly and the community guides stay fresher than any article.
- Get the official weights from Alibaba's published repositories. Verify you're on the real release, not a repack.
- Stand up a ComfyUI-style pipeline or the reference scripts, and start with a smaller Wan variant to confirm the plumbing works.
- Step up to the large variant your VRAM allows, adding quantization or offload only when the card forces it.
- Batch your generations. Slow per-clip time hurts least when the queue runs while you sleep.
If step 1 already sounds like a different hobby than "make a video," that's the signal to read the next section.
No GPU? The hosted math
Hosted Wan 2.5 costs money because someone else bought the GPU. On 8frame it's 102 credits per clip, which at pack rate (top-up packs run $10 per 1,000 credits and never expire) is about $1.02. There's no free tier on 8frame; plans start at $19 a month for 1,000 credits, which covers roughly 9 Wan 2.5 clips or 20 Wan 2.2 clips, watermark-free, with unused credits rolling over while you're subscribed.
The trade against self-hosting is clean. You give up "free forever" and get back the $1,000-class GPU you didn't buy, the weekend you didn't spend on setup, and clips in a couple of minutes in a browser tab. You also get what a local install can't offer: Wan sitting next to Veo, Kling, and the rest on one canvas, so the draft that proves a shot works can be finished on a stronger model without touching a second tool. The browser workflow is walked through in how to use Wan 2.5 online, and the prompt patterns that make Wan earn its keep are in Wan 2.5 prompts for free b-roll.
Which route fits you
| Your situation | Best route |
|---|---|
| You own a 24GB-class GPU and like owning pipelines | Self-host (genuinely free per clip) |
| You render overnight batches of b-roll | Self-host |
| You have no serious GPU | Hosted Wan 2.5, 102 credits on 8frame |
| You need clips today, not after a setup weekend | Hosted |
| You want Wan next to stronger models for finishing | 8frame canvas |
FAQ
Is Wan 2.5 actually free?
Not Wan 2.5, which has no public weights. Wan 2.2 does, under Apache 2.0, so running it on your own hardware costs nothing per clip, carries no watermark, and has no daily cap. The cost shows up as hardware instead: a 24GB-VRAM-class GPU is the practical minimum for full-quality output.
What GPU do I need to run Wan 2.5 locally?
A 24GB-class card is the realistic floor for the large, full-quality variants, which want up to 22 to 26 GB of VRAM at full precision. Smaller Wan variants run on 8 to 12 GB, or even 4 to 6 GB, with visible quality and resolution trade-offs. Quantization and CPU offload stretch smaller cards further at the cost of speed.
What does hosted Wan 2.5 cost if I skip the hardware?
On 8frame, 102 credits per clip, about $1.02 at pack rate, with no GPU, setup, or watermark involved. The $19 Starter plan's 1,000 monthly credits cover roughly 9 Wan 2.5 clips, and since all major models share one balance, winning drafts can be finished on a stronger model from the same canvas.
If the GPU math doesn't work for you, the hosted math is at 8frame pricing: about $0.65 a clip for watermark-free Wan 2.5, in the browser, next to every other major model.