Wan AI is Alibaba's video generation model series, built for text-to-video at low cost, with Wan 2.2 as the last release that shipped downloadable weights and Wan 2.5 and 3.0 available through APIs only.
The open branch of the family stops at Wan 2.2, published under Apache 2.0 on Hugging Face, which anyone can run locally. That's what separated Wan from closed models like Veo 3 or Sora 2. Wan 2.5 changed the pattern: it shipped as a hosted API with no public weights, and Wan 3.0 followed. On 8frame, Wan 2.2 runs 50 credits per 5-second 720p clip and Wan 2.5 runs 102 per 5-second 1080p clip, about 20 or 9 clips on the $19/month Starter plan, watermark-free. It outputs 720p at 24fps with generation times around 45 seconds per 5-second clip. The tradeoff for that price is real: it's a capable model for b-roll, prototyping, and abstract content, but the resolution floor and lighting depth sit below what closed paid-tier models produce.
How Wan AI works
Wan 2.5 is a video diffusion model. You give it a text prompt describing a scene, subject, camera motion, and lighting, and the model generates a video by progressively denoising a latent noise signal conditioned on your prompt.
For Wan 2.2, the weights are publicly released, so developers can run it on their own hardware. For Wan 2.5 and 3.0, access is through a hosted API. On 8frame, both run on managed infrastructure so you don't need a local GPU. The workflow is the same as any other model on the canvas: write a prompt, pick Wan 2.5 from the model picker, set your aspect ratio, and generate.
Prompt structure matters more for Wan 2.5 than for higher-parameter models. The model's defaults for lighting and background fill are generic, so you need to be explicit: name a light direction, a color temperature, a camera move, and a duration. A vague prompt produces flat output. A specific one produces something usable.
When you use Wan AI
Wan 2.5 earns its place in three situations.
Budget-constrained projects. At 50 credits per clip, Wan 2.2 is one of the cheapest video options on 8frame, and self-hosting its Apache 2.0 weights on your own GPU drops the per-clip cost to zero. Student projects, indie productions, and YouTube channels where 720p is fine and iteration count matters more than peak quality are the core use case.
Storyboard-phase iteration. Before spending 112 credits on a full Veo 3.1 clip, run the composition through Wan 2.5 twice. You'll know whether the camera angle and lighting concept work before committing the credit. Once the concept is confirmed, switch to the higher-quality model for the final output.
Abstract and texture content. Water surfaces, sky and cloud movement, and transition overlays don't expose the resolution floor the way faces and fine architecture do. Wan 2.5 handles abstract b-roll close to paid-tier quality at a fraction of the cost.
Skip Wan 2.5 and move to Kling 3.0 or Veo 3.1 when the clip is going to a client, will play full-screen above 1080p, or requires precise lighting and fine detail.
Examples on 8frame
Abstract water texture: Prompt: "Extreme close-up of rippling water surface with light from above refracting into shifting gold caustic patterns. Camera static, straight down. Warm 4500K light source directly overhead. Gentle continuous ripple motion. 5 seconds, loop-friendly." Generation time: 42 seconds. Cost on 8frame: 102 credits on Wan 2.5, or 50 on Wan 2.2. The caustic light patterns shifted organically and the loop point was clean enough for use as a title-card overlay. This is where Wan 2.5 performs closest to paid-tier output.
Nature establishing shot: Prompt: "Tall pine forest on a misty morning. Dense fog sits low between tree trunks. Dappled early light filtering through the canopy from directly above, cool 4000K. Camera very slowly pans right. 16:9 horizontal, 7 seconds." The fog rendered with convincing volumetric behavior and the clip looped cleanly. Cost: 102 credits on Wan 2.5, billed in 5-second tiers. Usable as background b-roll for documentary or travel content.
Both prompts are drawn from tested output on the 8frame canvas. For eight additional prompts across urban scenes, interiors, travel, and sky, see Wan 2.5 prompts for free b-roll.
Related concepts
Wan 2.5 prompts for free b-roll covers eight tested prompt templates, the Subject + Environment + Light + Camera motion + Duration formula, and the model's failure modes in detail. If you're new to Wan 2.5, start there.
Best AI video generator 2026 puts Wan 2.5 alongside Veo 3.1, Sora 2, Kling 3.0, and Seedance 2.0 in a direct model comparison. If you're deciding which model fits your production type, that's where to look.
What is b-roll? covers the broader concept Wan 2.5 is most commonly used to generate.
Ready to run Wan 2.5 without buying a GPU? Open 8frame and select Wan 2.5 from the model picker. Plans start at $19/month for 1,000 credits, about 9 Wan 2.5 clips or 20 Wan 2.2 clips, with no watermark on any plan.