Search for Wan 2.7 open source and you will find confident pages pointing to downloads, hardware requirements, and an Apache license. The problem is that confidence is not a release artifact. Before installing a large checkpoint or planning a production pipeline, verify the model against the publisher's repositories, model cards, license, and hosted documentation.
Why the Wan 2.7 open-source answer is confusing
Wan is both a family name and a collection of products. Alibaba can expose a model through its website or cloud API without releasing the weights required to run that exact version locally. A public API reference is evidence that the model exists. It is not evidence that its checkpoint, training code, or license is downloadable.
Search results often blur three different claims: a hosted model is available, a public repository exists, and the model weights are licensed for local use. Those claims require different proof. A page that links to an API and then supplies an unrelated community checkpoint has not established that the checkpoint is official.
This guide uses the official surfaces visible on September 2, 2026. Release status can change, so use the links below as a verification method rather than treating any article, including this one, as permanent.
What the official sources show today
| Release or route | Official evidence | What you can safely conclude |
|---|---|---|
| Wan 2.7 hosted services | Alibaba Cloud Model Studio lists text-to-video, image-to-video, reference-to-video, and video-editing model IDs | Wan 2.7 can be used through documented hosted routes where the model and region are available |
| Wan 2.7 local weights | No Wan 2.7 model or repository appears in the official Wan-AI Hugging Face and Wan-Video GitHub listings checked | An official downloadable Wan 2.7 release was not verified |
| Wan 2.2 local weights | The official repository links model downloads, setup commands, tasks, and an Apache 2.0 license | Wan 2.2 has a verifiable local inference path |
| Community files named Wan 2.7 | Names and mirrors vary, often without an official backlink or publisher checksum | The filename alone does not prove origin, model identity, or license |
Start with Alibaba's current Wan video-generation documentation. It lists Wan 2.7 hosted model IDs and separates international, global, US, and Chinese-mainland availability. The exact region matters because a model, endpoint, and API key must be compatible.
Then check the official Wan-AI organization on Hugging Face and the official Wan-Video repository list. At the time checked, the numbered open collections and repositories visible there stop at Wan 2.2.
Wan 2.7 is real, but hosted access is the documented path
The lack of an official local release does not mean Wan 2.7 is imaginary. Alibaba's documentation describes a substantial hosted family. Text-to-video supports audio, multi-shot narratives, 720p or 1080p output, and durations from 2 to 15 seconds in the listed international routes. Image-to-video adds first-frame, first-and-last-frame, and continuation workflows. Reference-to-video can use multiple entities, while the video-editing route accepts multimodal instructions.
Those capabilities vary by route and region. For example, the global and US tables can recommend Wan 2.6 while the international Singapore table lists Wan 2.7. Read the row for the model ID you will actually call rather than carrying a feature from one row into another.
Hosted use also has a different cost and privacy model than local inference. Review the current Alibaba Cloud model-pricing table, data-region rules, input requirements, and terms before uploading client assets. Prices and free quotas can change, so this guide does not freeze a per-second figure into the article.
Wan 2.2 is the verifiable open local option
If your real requirement is local Wan inference, use a release you can trace. The official Wan2.2 repository provides setup instructions and official Hugging Face and ModelScope links for text-to-video, image-to-video, text-image-to-video, speech-to-video, and animation variants.
Local does not automatically mean cheap or easy. The repository says its 5B text-image-to-video path can run with at least 24 GB of VRAM using offloading and reduced-memory options. The A14B text-to-video and image-to-video examples call for at least 80 GB of VRAM. Your generation time, electricity, storage, setup labor, and failed runs all belong in the cost calculation.
The repository itself is Apache-2.0 licensed, but do not reduce a commercial review to one badge. Confirm the license files attached to the exact weights and dependencies you deploy, then review the service, client, and output requirements that apply to your use case.
Use this five-point download verification check
- Publisher: Is the file inside the official Wan-AI or Wan-Video organization, or linked directly from it?
- Release trail: Does an official announcement or repository changelog name the exact version?
- License: Is there a license attached to the exact model artifact, not just nearby code?
- Integrity: Does the publisher provide a checksum, commit, revision, or immutable model-card history you can verify?
- Reproducibility: Do the official inference instructions name the same architecture, filenames, and dependencies as the download?
If one of those checks fails, do not treat the file as an official Wan 2.7 release. A community fine-tune or renamed checkpoint may still be useful, but it requires its own provenance, safety, and license assessment. Never run an unknown installer or model bundle just because a search result calls it official.
Choose local, API, or a browser workflow by the job
Choose local Wan 2.2 when
You need a verifiable open release, control the infrastructure, accept setup and GPU work, and can test the exact licenses and dependencies.
Choose the Wan 2.7 API when
You need the documented Wan 2.7 capabilities, can work within available regions, and prefer metered hosted inference over maintaining GPUs.
Choose a browser studio when
Your real goal is finishing clips, comparing available models, and connecting video with images, voice, music, or editing without managing infrastructure.
QuestStudio fits the third path. It is not presented here as a Wan 2.7 host. Its value is a browser-based, multi-model creative workflow for people who care more about approved output than owning a specific checkpoint. You can review QuestStudio's current model comparisons before choosing a test.
Compare cost per approved clip, not access price
A local model can avoid per-call fees but still lose economically if setup, generation time, maintenance, and failed outputs consume more resources than hosted inference. An API can be fast to integrate but become expensive when every acceptable clip requires multiple retries. A subscription can look simple while hiding model-specific credit costs.
Run one representative five-second shot. Record hardware or cloud cost, elapsed time, attempts, output resolution, and whether the clip passes your acceptance criteria. Divide the total cost by approved clips. The AI generation cost calculator can keep local, API, and studio workflows in the same comparison.
Use the same source image, aspect ratio, motion requirement, and retry cap. If one route supports audio and another does not, count the extra audio workflow rather than pretending the outputs are equivalent.
A practical 20-minute decision test
- Name the requirement. Decide whether you need Wan 2.7 specifically, local control, a certain motion result, or simply a usable video.
- Verify the release surface. Check Alibaba Cloud, the official Hugging Face organization, and the official GitHub organization on the same day.
- Prepare one shot. Use a five-second brief with a subject, action, camera move, aspect ratio, and one pass/fail rule.
- Test two viable routes. Compare only routes you can access safely and legally. Keep the prompt and retry cap stable.
- Choose by finished output. Score adherence, identity, motion, artifacts, recovery time, and cost per approved clip.
The AI video model benchmark workflow expands that test when you need a repeatable team evaluation. If a browser workflow is the better fit, open QuestStudio Video Lab with article attribution and test a currently available model against the same acceptance criteria.
Wan 2.7 open-source FAQ
Is Wan 2.7 open source?
As of September 2, 2026, Alibaba documents Wan 2.7 through hosted Model Studio services, while the official Wan-AI Hugging Face and Wan-Video GitHub organizations do not list official Wan 2.7 weights or a Wan 2.7 repository. Treat third-party download claims as unverified unless an official release links to them.
Can I download Wan 2.7 and run it locally?
No official local Wan 2.7 package was verified in the official release surfaces checked for this guide. Do not assume a third-party file is official because it uses the Wan name. Verify the publisher, license, checksum, and official backlink first.
What is the latest officially downloadable Wan video release?
Wan 2.2 is the latest numbered video release visible in the official Wan-AI model collections and Wan-Video repository list checked on September 2, 2026. Its official repository provides model links, installation instructions, and an Apache 2.0 license.
How can I use Wan 2.7?
Alibaba Cloud documents Wan 2.7 text-to-video, image-to-video, reference-to-video, and video-editing services through Model Studio. Availability depends on the model, region, endpoint, and account configuration.
Do I need to self-host an AI video model?
Self-hosting is useful when local control, infrastructure ownership, or a specific open model is the real requirement. Creators who mainly want to compare outputs and finish clips may be better served by a hosted API or browser-based multi-model workflow.
The safest answer is a dated, verifiable answer
Wan 2.7 has documented hosted capabilities. Wan 2.2 has a documented open local path. The official release surfaces checked for this guide do not currently connect Wan 2.7 to downloadable official weights. That distinction protects your time, hardware budget, client data, and production plan.
If Alibaba publishes official Wan 2.7 weights later, the correct answer should change with the evidence. Until then, choose a route you can verify and judge it by the work it finishes.