Wan 2.5 Review — What It Is, Is It Open Source & How Good Is It Really
Updated July 2026
Wan 2.5 is a generation of Alibaba's Tongyi Wanxiang (通义万相) video model family, announced at the September 2025 Yunqi Conference and commercially launched that November. It shipped as four models — text-to-video, image-to-video, text-to-image and image editing — and its defining feature was native synchronized audio: dialogue, effects and ambience generated with the picture rather than dubbed on after, in clips of up to 10 seconds at up to 1080p.
One thing to clear up immediately, because it's the most common misconception about this model: Wan 2.5 is not open source. Wan earned its open-weights reputation on releases 2.1 and 2.2, genuinely published under Apache 2.0. Wan 2.5 is API-only, and its weights have never appeared on the official Wan-AI Hugging Face organisation. This review covers what the model does, how good it honestly is against Kling 3.0 and Veo 3.1, and what it costs.
| Developer | Alibaba (Tongyi Lab) |
|---|---|
| Announced | Sept 2025; commercial launch Nov 2025 |
| Models | T2V, I2V, T2I, image-edit — all still '-Preview' |
| Output | Up to 1080p, 5 or 10s, native synchronized audio |
| Open weights | No — the Apache 2.0 line stops at Wan 2.2 |
| Official site | wan.video · Alibaba Cloud Model Studio |
What Wan 2.5 costs (published rates, July 2026)
| Plan | Price | What you get |
|---|---|---|
| 480p (via fal) | $0.05/second | $0.25 per 5s clip, $0.50 per 10sPublished rate, July 2026 |
| 720p (via fal) | $0.10/second | $0.50 per 5s clip, $1.00 per 10s |
| 1080p (via fal) | $0.15/second | $0.75 per 5s clip, $1.50 per 10s |
| Alibaba Cloud Model Studio | ~$0.105/second | First-party DashScope APIReported — no public video price table |
| wan.video (consumer app) | Credit packs & subscriptions | Browser access to the hosted modelVaries by region and promo |
| Self-hosting Wan 2.1 / 2.2 | $0 licence + GPU | Apache 2.0 weights on GitHub and Hugging FaceThe open versions — not 2.5 |
The per-second rates above are fal's published price card for the Wan 2.5 endpoints, captured July 2026 — posted rates, not estimates, and in line with what other hosts charge. Alibaba's own Model Studio documentation exposes no public per-second table for video, so the DashScope figure is secondary and marked as such. Note the shape: resolution is the only real lever and it scales linearly, which makes Wan 2.5 unusually easy to budget.
The honest read
The open-source question, settled
This deserves a straight answer, because a lot of published writing gets it wrong. Wan built its name on genuinely open releases: Wan 2.1 in February 2025 and Wan 2.2 in mid-2025, both Apache 2.0, both with downloadable weights. Check the official Wan-AI organisation on Hugging Face today and what's published is still that 2.1/2.2 line — T2V, I2V and TI2V checkpoints plus derivatives like Wan2.2-Animate, Wan2.2-S2V and the newer Wan-Dancer. There is no Wan 2.5 repository, and there never was; Alibaba moved to an API-first commercial model from 2.5 onward. So if you want Wan on your own GPU, in ComfyUI, or fine-tuned, you're running 2.2. If you want 2.5, you're renting it.
What 2.5 actually brought: audio as a first-class output
Strip away the version-number noise and 2.5's contribution is sound. Alibaba's launch material is explicit that the video models generate audio synchronized with the visuals rather than as a separate lip-sync pass — ahead of the field at the time, when most competitors still shipped silent clips. Paired with 1080p and prompt adherence that holds on multi-element prompts, that made Wan 2.5 unusually complete for sound-on social video. The limits are equally clear: 10 seconds per generation, so anything longer is a stitching job — and a model still carrying '-Preview' in its own identifier months after commercial launch.
Where it loses, honestly
Wan 2.5 is not the quality leader and it's worth saying plainly. On photorealism and cinematic polish — faces especially, and fine texture — both Veo 3.1 and Kling 3.0 edge it out, and the gap widens on visually complex scenes where 2.5's coverage looks thinner. Kling is stronger on heavy motion and action choreography; Veo is more reliable when you need exactly the framing you described. What Wan 2.5 wins is the ratio: audio-synced 1080p at a fraction of what comparable engines charge per second. That's a value argument, not a quality crown.
Is 2.5 still the version to care about?
Alibaba has shipped further releases since — a 2.6 series, and a 2.7 line adding a reasoning step before generation — and their open-weights status is disputed enough across sources that we won't assert it either way. But 2.5 remains the version most widely deployed across third-party hosts and sits at the cheapest published tier of the family, which is why the bare search term stays busy. The practical question isn't 2.5 versus 2.7; it's whether Wan's audio-plus-price trade beats Kling's motion or Veo's fidelity for your shots — answered by running one prompt through all three engines.
Verdict
Use Wan 2.5 if you're making sound-on video at volume and budget matters. Native synchronized audio at 1080p for around $0.15 a second is the best price-per-finished-clip in its class, and linear per-second pricing makes costs trivial to forecast — no credits, no exchange rate, no expiry.
Skip it if you need the top of the quality range, or if you specifically wanted the open model. Veo 3.1 and Kling 3.0 are visibly better on faces, fine detail and heavy motion. And if your plan was to download weights and run locally, 2.5 won't do that — Wan 2.2 under Apache 2.0 is the model you actually want.
Run Wan 2.5 without the API setup
You don't need an Alibaba Cloud account or a GPU to use this class of model — Wan 2.5 runs here in the browser, first videos free, no card required. Weighing it against the motion-first option? Run the same prompt on Kling 3.0 and judge the two side by side. Finished cuts can go through a 4K upscaler before delivery.
FAQ
What is Wan 2.5?
A generation of Alibaba's Tongyi Wanxiang video model family, announced September 2025 and commercially launched that November. It generates video up to 10 seconds at up to 1080p from text or stills, with natively synchronized audio.
Is Wan 2.5 open source?
No. Wan's open-weights reputation belongs to Wan 2.1 and 2.2, published under Apache 2.0. Wan 2.5 is API-only — no weights have ever appeared on the official Wan-AI Hugging Face organisation.
Can I run Wan 2.5 locally?
Not the 2.5 models themselves. To run Wan on your own hardware or fine-tune it, you're running Wan 2.1 or 2.2 — those have downloadable checkpoints and mature ComfyUI workflows, and need a high-end GPU.
How much does Wan 2.5 cost?
On fal's published card it runs $0.05 per second at 480p, $0.10 at 720p and $0.15 at 1080p — about $1.50 for a 10-second 1080p clip with audio. Alibaba's DashScope API is reported near $0.105 per second.
Does Wan 2.5 generate audio?
Yes, and it's the model's defining feature. Audio is generated synchronized with the visuals rather than added as a separate lip-sync pass, covering dialogue, effects and ambience.
Wan 2.5 vs Kling 3.0 — which is better?
Kling 3.0 is stronger on motion, action choreography and photorealistic detail. Wan 2.5 is stronger on price and competitive on prompt adherence and audio sync. Kinetic shots, take Kling; sound-on clips at volume on a budget, take Wan.
What's the difference between Wan 2.2 and Wan 2.5?
Wan 2.5 adds natively synchronized audio, 1080p output and better prompt adherence — but is API-only. Wan 2.2 is the open Apache 2.0 release you can self-host, and its Animate variant remains the pick for character animation from reference video.