Correction, 2026-10-02: we said there is no free way to run Wan 2.5. Alibaba Cloud Model Studio gives new users a free quota of 50 seconds of output per Wan 2.5 model (Singapore region, valid 90 days), so the page now covers that trial and Alibaba's own per-second price after it. We said Alibaba stopped publishing Wan weights after 2.1 and 2.2 and called Wan "the only genuinely free route": Wan-AI has published Wan 2.2-based Animate models and a dance model since, and Wan is not the only open-weights video family, so both lines now say what we checked. The 8frame price now says it is for a 5-second clip at 1080p, and that Wan 3.0, the model the Wan tool opens on, costs less. We removed setup-time estimates we never measured ("an afternoon", "a weekend"), an unsourced claim that community quantization lowers the hardware bar at a cost in speed, and lines calling other models "stronger". Kling's free-plan terms are now attributed to its plan table, which a Kling FAQ contradicts on commercial use.
Correction, 2026-10-01: we gave VRAM figures for Wan variants (4 to 6 GB, 8 to 12 GB, 22 to 26 GB) that we cannot trace to a source, said a hosted Wan 2.5 clip arrives in a couple of minutes (we have no measured generation time for it) and put a price on the GPU you would need. The hardware section now quotes Wan 2.2's own model card, and the timing and GPU price are gone. The Wan 2.5 release date had no source; it now cites Alibaba Cloud's own Model Studio release list.
Correction, 2026-09-30: we said Pika watermarks free output and gave Kling "three generations a day", a figure Kling does not publish. Both now follow the companies' own pages.
There is no free-forever 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 the Wan 2.2 line, but not for the generations after it: Wan 2.5 arrived in September 2025 as a preview API on Alibaba Cloud Model Studio (wan2.5-t2v-preview and wan2.5-i2v-preview, listed for the international region on 2025-09-23) with no public weights, and Wan 2.6, 2.7 and 3.0 are also API models on Model Studio. We checked the Wan-AI organization on Hugging Face on 2026-10-02: its newest uploads, from July and August 2026, are Wan 2.2 Animate models and a music-to-dance model, Wan-Dancer-14B, and there are no Wan 2.5, 2.6, 2.7 or 3.0 weights. What Alibaba does give away is a trial: new Model Studio users get 50 seconds of free output per Wan 2.5 model. So if "wan 2.5 free" means a few test clips, that quota is the route. If it means free for good, the model you can actually download is Wan 2.2. This guide covers both, what self-hosting costs in hardware, and the hosted math if you need Wan 2.5 after the trial.
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 only free Wan 2.5 we could confirm is Alibaba Cloud Model Studio's new-user quota: 50 seconds of output per Wan 2.5 model, Singapore region, valid 90 days, then pay-as-you-go.
- The real price of self-hosting is hardware: Wan 2.2's model card runs its smaller TI2V-5B model on a GPU with at least 24GB of VRAM, plus real setup time.
- Free tiers of closed video tools come with catches: Kling's and Hailuo's free output is watermarked, Pika's free plan has no monthly credits, and all three keep commercial use for paid plans in their plan pages or policies.
- No GPU? Wan 2.5 on 8frame is 102 credits for a 5-second 1080p clip, about $1.02 at pack rate, roughly 9 clips on the $19 Starter plan. Wan 3.0 costs less (95 credits for 5 seconds at 1080p, 24 at 480p).
Why open weights are the genuinely free route
Free AI video usually means a demo. Kling's plan table lists no monthly credits on its free Basic plan, and its video generator page says free generations include a watermark; Hailuo's payment policy says free users' downloads carry a watermark; Pika's free plan has no watermark but includes 0 monthly credits and no commercial licence. On commercial use, Pika's pricing page leaves it out of the free plan, Hailuo's policy gives you the rights to content made on a paid plan, and Kling's plan table says Basic content "is not for commercial use" (a FAQ on Kling's video generator page says the opposite; go by the plan table). All of this is from each company's own pages, checked 2026-10-02. Those tiers exist to let you evaluate a closed model before paying, and they're fine at that job, but none of them hands you a clip you are cleared to ship commercially.
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 free-tier watermark because there's no free tier: no company sits between you and the model. No daily cap, because the compute is yours. And the Wan 2.2 model card says Alibaba claims "no rights over the your generated contents" (their wording), within the limits of its Apache 2.0 licence. If "free" in your search meant "free for good, and I can use the output," self-hosted Wan 2.2 is the honest answer. It isn't the only open-weights video model, and each comes with its own licence terms: the best open-source AI video generators of 2026 compares them, 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 a 50-second trial is enough, or whether you own the machine Wan 2.2 takes.
The free trial: Alibaba's new-user quota
Alibaba Cloud Model Studio, the API platform where Wan 2.5 lives, grants new users a free quota per model. Its pricing page, checked 2026-10-02, lists 50 seconds of free output for wan2.5-t2v-preview (text to video) and another 50 seconds for wan2.5-i2v-preview (image to video). The rules, from Alibaba's own free-quota page:
- Singapore region only. The quota applies to models in the Singapore region with the International deployment scope; other regions get none.
- New users only. It is granted when you first activate Model Studio, after you complete your account information, and it can take up to two hours to show up. Re-registering a new account does not grant another one.
- 90 days. The quota is valid for 90 days from activation (or the model's release or your access approval, whichever is later), and unused seconds expire.
- Per model. Each model has its own quota, so text-to-video and image-to-video each get their 50 seconds; it doesn't move between models.
- Billing starts after it. When the quota runs out you are billed pay-as-you-go unless you turn on "Free Quota Only", which is off by default and stops the model instead. Switch it on before your first call if you want the trial to stay free.
This is a developer platform: the quota is spent on API calls made with your Alibaba Cloud API key. After the quota, Model Studio's own price for Wan 2.5 in that region is $0.05 per second at 480P, $0.10 at 720P and $0.15 at 1080P, so a 5-second 1080P clip costs $0.75. That's the one free Wan 2.5 offer we could confirm on an Alibaba page on 2026-10-02; if you see another, check it on Alibaba's own page before you count on it.
What self-hosting actually takes
If you want Wan for free for good, here's the entry ticket for Wan 2.2, stated plainly.
The GPU. Wan 2.2's main generation models are T2V-A14B and I2V-A14B (480P and 720P) and the smaller TI2V-5B (720P, text and image to video); the line also has an audio-driven model (S2V-14B) and Animate models. The TI2V-5B model card says its generation command "can run on a GPU with at least 24GB VRAM (e.g, RTX 4090 GPU)", with model offloading on; on a GPU with at least 80GB, the card says you can switch the offloading off to speed things up. If your card has less memory, look for community quantized builds and check the requirements each one states; the official card doesn't cover them.
The setup. The TI2V-5B repository on Hugging Face holds about 34 GB of files and T2V-A14B about 126 GB. Then 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. Set aside real time for it before you count on the first clip.
The wait. The same model card says TI2V-5B makes "a 5-second 720P video in under 9 minutes on a single consumer-grade GPU". 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 can be a good deal: no per-clip fee, only electricity and your time.
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 the Wan-AI organization on Hugging Face (or its ModelScope mirror). Verify you're on the real release, not a repack.
- Stand up a ComfyUI-style pipeline or the reference scripts, and start with TI2V-5B to confirm the plumbing works.
- Step up to an A14B model if 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, Wan 2.5 makes 5-second clips at 1080p for 102 credits each, which at pack rate (top-up packs run $10 per 1,000 credits) is about $1.02. 8frame has no free trial and signup includes no credits; plans start at $19 a month for 1,000 credits, which covers roughly 9 Wan 2.5 clips, watermark-free, with unused credits rolling over while you're subscribed. Per clip, that is more than Alibaba's own $0.75 after the trial; what you pay for is everything around the model.
If you don't need Wan 2.5 specifically, Wan 3.0 costs less on 8frame: 24, 48 or 95 credits for 5 seconds at 480p, 720p or 1080p, with clips from 2 to 30 seconds. The Wan tool opens on Wan 3.0 at 1080p, so switch the model to 2.5 if you want Wan 2.5 itself; Wan 3.0 vs Wan 2.5 compares the two. On Starter, 1,000 credits is 10 Wan 3.0 clips at 1080p or 41 at 480p. Wan 2.2 is there too at 50 credits per 5 seconds at 720p (20 clips), but on 8frame it needs a start image or a video; it has no text-to-video.
The trade against self-hosting is clean. You give up "free forever" and get back the GPU you didn't buy, the setup time you didn't spend, and clips in a browser tab. You also get what a local Wan install doesn't give you: Wan next to Veo, Kling, Seedance and the rest on one canvas and one credit balance, so a shot you drafted on Wan can be re-run on another model without a second tool. The browser workflow is walked through in how to use Wan 2.5 online, and prompt patterns for Wan are in Wan 2.5 prompts for free b-roll.
Which route fits you
| Your situation | Best route |
|---|---|
| You want to try Wan 2.5 itself without paying | Alibaba Cloud Model Studio's new-user quota (50 seconds per Wan 2.5 model, 90 days) |
| You own a GPU with 24GB of VRAM or more and like owning pipelines | Self-host Wan 2.2 (no per-clip fee) |
| You render overnight batches of b-roll | Self-host Wan 2.2 |
| You have no serious GPU | Hosted: Wan 2.5 at 102 credits for 5 seconds on 8frame, or Wan 3.0 from 24 |
| You need clips today, not after a setup project | Hosted |
| You want Wan next to other models on one canvas | 8frame canvas |
FAQ
Is Wan 2.5 actually free?
Not for good. Wan 2.5 has no public weights, so you can't run it on your own hardware. The free access we could confirm is a trial: Alibaba Cloud Model Studio gives new users 50 seconds of output per Wan 2.5 model, in the Singapore region, valid 90 days. Wan 2.2 does have open weights, under Apache 2.0, so running it on your own hardware costs nothing per clip, carries no free-tier watermark, and has no daily cap. The cost shows up as hardware instead: Wan 2.2's model card runs TI2V-5B on a GPU with at least 24GB of VRAM.
How do I get Alibaba's free Wan 2.5 quota?
Create an Alibaba Cloud account, complete your account information, and activate Model Studio in the Singapore region. The quota is granted automatically on first activation (it can take up to two hours to appear): 50 seconds for wan2.5-t2v-preview and 50 for wan2.5-i2v-preview, for 90 days. You use it through the API with your API key. Turn on "Free Quota Only" first, because it is off by default and without it you are billed pay-as-you-go once the quota runs out.
What GPU do I need to run Wan 2.5 locally?
Wan 2.5 cannot run locally: it has no published weights. For Wan 2.2, the model card for TI2V-5B, the smaller model in the line, names a GPU with at least 24GB of VRAM (an RTX 4090, for example) with offloading on, and 80GB to drop the offloading. For smaller cards, check what community quantized builds say they need.
What does hosted Wan 2.5 cost if I skip the hardware?
On 8frame, 102 credits for a 5-second clip at 1080p, 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, or 10 Wan 3.0 clips at 1080p, and since every model on 8frame draws on the same credit balance, a draft can be re-run on another model from the same canvas. On Alibaba Cloud Model Studio, after the free quota, Wan 2.5 costs $0.15 per second at 1080P.
Sources
- Kling free plan and watermark: kling.ai/feature/ai-video-generator (free generations include a watermark; its FAQ also says videos can be used commercially) and the plan table in its Kling 3.0 credit guide (Basic: no monthly credits, "generated content is not for commercial use"), checked 2026-10-02
- Hailuo watermark and content rights: hailuoai.video/doc/payment-policy.html, checked 2026-10-02
- Pika: pika.art/pricing, checked 2026-10-02
- Wan 2.2 models, licence, rights over generated content, VRAM and generation time: huggingface.co/Wan-AI/Wan2.2-TI2V-5B, checked 2026-10-02
- Wan-AI published repositories (newest uploads, no Wan 2.5, 2.6, 2.7 or 3.0 weights) and repository sizes: huggingface.co/Wan-AI, read through the Hugging Face models API, checked 2026-10-02
- Wan 2.5 preview release on Model Studio (wan2.5-t2v-preview, wan2.5-i2v-preview, 2025-09-23, international) and Wan 3.0 as a Model Studio model (2026-08-06): Alibaba Cloud, alibabacloud.com/help/en/model-studio/newly-released-models, checked 2026-10-02
- Wan 2.5, 2.6 and 2.7 on Model Studio, Wan 2.5 per-second prices and the 50-second free quota: Alibaba Cloud, alibabacloud.com/help/en/model-studio/model-pricing, checked 2026-10-02
- Free quota rules (Singapore only, account information, 90 days, per model, no new quota on re-registering, Free Quota Only off by default): Alibaba Cloud, alibabacloud.com/help/en/model-studio/new-free-quota, checked 2026-10-02
- 8frame prices (Wan 2.5, Wan 3.0, Wan 2.2), Wan 2.5 clip length, Wan tool defaults and Wan 2.2 inputs: 8frame tool registry and content fact sheet, read 2026-10-02
If the GPU math doesn't work for you, the hosted math is at 8frame pricing: about $1.02 for a 5-second Wan 2.5 clip at 1080p, watermark-free, in the browser, next to Veo, Kling and Seedance.
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.