Risposta rapida: Wan 2.5 is Alibaba’s hosted AI video model for text-to-video and image-to-video generation. Its official Model Studio endpoints can produce 5- or 10-second MP4 clips at 480P, 720P, or 1080P, with synchronized audio generated alongside the visuals. The model remains in Alibaba Cloud’s documentation, although it is no longer the newest release in the Wan family.
Much of the confusion around Wan 2.5 comes from its relationship to Alibaba’s open-source Wan models. They belong to the same family, but the official Wan 2.5 video endpoints were released as hosted Preview services. Alibaba has not published verified Wan 2.5 weights for local ComfyUI use, so access, licensing, and setup differ from open-weight releases such as Wan 2.2.
For readers who prefer keeping chat, research, image, video, and agent tools in one place, GlobalGPT provides an all-in-one workspace under one account. It can be a practical way to check whether Wan 2.5 or a newer Wan model is available without moving between several provider sites. Because availability changes, check the live Video picker before starting a project.
Indice
What Is Wan 2.5 AI?
Wan 2.5 is part of Alibaba’s Wan visual-generation family. Alibaba previewed four Wan 2.5 models at Apsara Conference 2025, including video generation, image generation, and image editing models. The company described the video models as using a jointly trained text, audio, and visual architecture designed to improve prompt following and synchronize sound with the generated picture.
The timing explains the original headline. In its September 2025 Wan 2.5 announcement, Alibaba positioned the model alongside its larger full-stack AI push. Wan 2.5 was not simply another silent video generator. Native audio was the feature that moved it into the same conversation as premium hosted video systems.
By September 2026, however, Alibaba’s product line had moved forward. The official Model Studio catalog still lists Wan 2.5 Preview, while newer Wan 2.7 and Wan 3.0 endpoints offer longer clips or broader reference controls. If you are researching the current family rather than this specific version, start with the Wan 3.0 release and feature guide.
Wan 2.5 Specifications
Il Alibaba Cloud Model Studio catalog, updated September 2, 2026, lists two Wan 2.5 Preview endpoints:
| ID modello | Input mode | Audio | Risoluzione | Durata | Uscita |
|---|---|---|---|---|---|
wan2.5-t2v-preview | Da testo a video | Audio sync | 480P, 720P, 1080P | 5 or 10 seconds | 30 fps, MP4 |
wan2.5-i2v-preview | Da immagine a video | Audio sync | 480P, 720P, 1080P | 5 or 10 seconds | 30 fps, MP4 |
La parola Anteprima is not decorative. It is part of both model IDs and signals a hosted product endpoint, not a downloadable checkpoint. It also means you should verify access and billing in the platform you plan to use instead of trusting an old bookmarked route.
What Can the Wan 2.5 AI Video Generator Do?
Generate video and sound together
Wan 2.5’s defining upgrade was synchronized audio. Alibaba said the model could create high-fidelity audio with the video rather than requiring creators to build the sound layer in a separate tool. For a short dialogue, product shot, or atmospheric scene, that can remove a full editing step.
Turn prompts into short 1080P clips
The text-to-video endpoint accepts a written scene description and returns a 5- or 10-second clip. A useful prompt should describe the subject, action, camera, setting, visual treatment, and sound. That structure is more reliable than stacking vague quality words such as “cinematic” and “stunning.”
Animate a still image
The image-to-video endpoint gives the model a visual starting point. This is useful when the exact product, character, composition, or color palette matters more than open-ended invention. For a broader view of the market, the Confronto tra i migliori generatori di video basati sull'intelligenza artificiale shows how different tools approach quality, control, and workflow.
Wan 2.5 Image to Video: A Practical Workflow
Start with a clean source image. A frame with one obvious subject, readable separation from the background, and enough room for movement gives the model fewer conflicts to solve. If the source already contains distorted hands, unreadable text, or impossible geometry, motion usually makes those flaws more visible.
- Choose the first frame. Use the image that best represents the subject and final art direction.
- Describe motion, not appearance alone. Tell the model what the subject does and how fast it happens.
- Direct the camera. Use one clear move such as a slow push-in, lateral tracking shot, or locked camera.
- Write the sound layer. Describe dialogue, ambience, sound effects, or music only when it helps the scene.
- Generate a short draft first. Fix motion and composition before spending more on higher resolution.
Image-to-video models do not all behave the same way. The Kling AI image-to-video workflow is a useful comparison when you need a second approach to source-image motion and social video production.
How to Use Wan 2.5 in 2026
Option 1: Alibaba Cloud Model Studio
Model Studio is the official hosted route for the two Wan 2.5 Preview endpoints. It is the clearest choice for developers who need named model IDs, API-controlled jobs, and provider documentation. You will need an Alibaba Cloud account, the correct deployment scope, and enough quota or billing balance for the job.
Option 2: GlobalGPT’s video workspace
GlobalGPT is designed for users who would rather keep several AI workflows under one account and one subscription. Its pagina del piano pubblico listed Wan 2.5 among its included video models when checked on September 3, 2026, alongside newer Wan options. The practical benefit is broader capability coverage without maintaining separate subscriptions for every provider.
Open the Video workspace and inspect the live model picker before starting. If Wan 2.5 is missing, do not force an old dedicated URL. Choose Wan 2.7 or Wan 3.0 based on the controls and duration you need. Users comparing the newest family can also review Wan 3.0 Prime’s speed and pricing differences.
Is Wan 2.5 Free?
Wan 2.5 is not a freely downloadable open model. It is a hosted Preview service, so usage depends on the provider’s current quota, credits, or billing rules. A new-user allowance or promotional credits should be treated as a limited trial, not as permanent free access.
This is another reason to compare the whole workflow rather than one headline price. Resolution, duration, failed generations, and audio can change the real cost per usable clip. For the newer model family, the Wan 3.0 pricing and API cost guide provides more current planning context.
Is Wan 2.5 Open Source?
No official Wan 2.5 model weights were listed in Alibaba’s public Wan repository when checked on September 3, 2026. Il sito ufficiale Wan2.2 GitHub repository provides inference code, checkpoints, and an Apache 2.0 license for Wan 2.2 releases. It does not turn the separately hosted Wan 2.5 Preview endpoints into open-source downloads.
This is the common source of confusion behind the keyword “Wan 2.5 open source.” Wan is an open model family in a broad sense, but every numbered product is not automatically released under the same terms. API access means you can call a model. Open weights mean you can download and run its parameters. Those are different things.
Can You Run Wan 2.5 in ComfyUI?
There is no verified official local Wan 2.5 ComfyUI workflow without official Wan 2.5 weights. Alibaba’s Wan2.2 repository explicitly records ComfyUI integration for Wan2.2 text-to-video, image-to-video, text-image-to-video, speech-to-video, and animation tasks. That evidence applies to Wan 2.2, not Wan 2.5 Preview.
A third-party ComfyUI node may still call a hosted Wan 2.5 API. That is an API wrapper, not local inference. Check where the node sends your prompt and source images, how it stores credentials, and how it charges for generations before installing it.
If local control matters more than this exact version number, use an officially released Wan 2.2 checkpoint and its documented ComfyUI workflow. The trade-off is that the core Wan 2.2 text-to-video and image-to-video endpoints do not match Wan 2.5 Preview’s native synchronized-audio feature set.
Wan 2.5 vs Newer Wan Models
| Versione | Il motivo principale per cui vale la pena usarlo | Durata e risoluzione | Local/open workflow |
|---|---|---|---|
| Wan 2.2 | Official checkpoints and ComfyUI support | Varies by checkpoint; core T2V/I2V supports 480P or 720P | Yes, for officially released Wan 2.2 models |
| Wan 2.5 Preview | Short video with synchronized audio | 5 or 10 seconds; 480P, 720P, 1080P | No official weights verified |
| Wan 2.7 | Longer hosted clips and multi-shot features | Up to 15 seconds; 720P or 1080P for T2V/I2V | Hosted endpoints in Model Studio |
| Wan 3.0 | Longer clips and broader reference inputs | 2 to 30 seconds; 480P, 720P, 1080P | Hosted endpoints in Model Studio |
Wan 2.5 still makes sense when you need its exact hosted endpoint or want to reproduce an older workflow. For new projects, compare it with the latest option first. The Wan 3.0 vs Kling 3.0 comparison is a better starting point for a current buying decision.
Wan 2.5 Prompt Checklist
Use this order to keep the prompt readable:
Subject + action + setting + camera + lighting/style + audio + constraints
Example text-to-video prompt: A ceramic coffee cup sits on a dark wooden counter as steam curls upward. The camera makes a slow push-in. Warm morning light enters from the left. Quiet cafe ambience, a soft cup clink, no dialogue, no on-screen text.
Example image-to-video prompt: Keep the product shape and label unchanged. Condensation slowly forms on the bottle while the camera tracks from left to right. Cool studio lighting, shallow depth of field, subtle ice crackle and room tone. No extra objects and no text distortion.
- Use one main action per short clip.
- Choose one camera move instead of combining several.
- Write exact audio cues rather than asking for “good sound.”
- Protect important product details with explicit constraints.
- Generate at a lower resolution while testing motion, then increase quality.
For adult or restricted-content questions, separate model capability from platform policy. The Wan content-policy and prompt workflow guide explains the practical limits without treating every generated request as acceptable.
Domande frequenti
What is Wan 2.5?
Wan 2.5 is an Alibaba AI video model released as a hosted Preview service. Its official text-to-video and image-to-video endpoints create 5- or 10-second MP4 clips at up to 1080P and include synchronized audio generation.
Is Wan 2.5 an Alibaba AI video model?
Yes. Wan is Alibaba’s visual-generation model family, and Alibaba introduced Wan2.5 Preview at Apsara Conference 2025. The company presented four Wan2.5 models covering video generation, image generation, and image editing.
Is Wan 2.5 open source?
No official Wan 2.5 weights were listed in Alibaba’s public Wan repository as of September 3, 2026. Open-source Wan 2.2 checkpoints exist, but that does not make the hosted Wan 2.5 Preview endpoints open source.
Can I use Wan 2.5 with ComfyUI?
Not as verified local inference without official Wan 2.5 weights. Some third-party nodes may call a hosted API, but that is different from running the model locally. Official ComfyUI integration is documented for released Wan 2.2 checkpoints.
Is Wan 2.5 free to use?
Wan 2.5 is not a permanently free downloadable model. Hosted access can use provider credits, trial quota, or paid billing. Check the live platform before generating because account eligibility, credit costs, and model availability can change.
Does Wan 2.5 support image-to-video generation?
Yes. The official wan2.5-i2v-preview endpoint turns an input image into a 5- or 10-second video. It supports 480P, 720P, and 1080P MP4 output at 30 fps with synchronized audio.
Should I still use Wan 2.5 in 2026?
Use Wan 2.5 when you need that exact Preview endpoint or must reproduce an older workflow. For a new project, compare Wan 2.7 and Wan 3.0 first because Alibaba’s current catalog gives them longer duration or broader controls.
Verdetto finale
Wan 2.5 was an important step in Alibaba’s AI video push because it combined text-to-video and image-to-video generation with synchronized sound. Its official specifications remain useful: two hosted Preview endpoints, up to 1080P, 5- or 10-second duration, 30 fps, and MP4 output.
