Is Wan 2.7 Open Source? What You Can Actually Download

Is Wan 2.7 open source? No. As of July 20, 2026, the official Wan organization's open weights stop at Wan 2.2 — the T2V-14B, I2V-14B, and TI2V-5B models on Hugging Face and GitHub. Sites offering "Wan 2.7" downloads, weights, or an "Apache 2.0 release" are SEO pages with no official release behind them. The newest real Wan release is Wan-Dancer-14B (July 13, 2026) — and per its release materials it targets datacenter hardware, not your desktop.

If you searched "wan 2.7 download" and landed on a page that looked official, you're not alone. This guide covers what actually exists, why the fake version numbers keep appearing, and — more usefully — exactly which Wan model fits your GPU today.

---

The Short Answer {#the-short-answer}

Two things follow from that table. First, if a page tells you Wan 2.7 is "open source under Apache 2.0" with "weights on Hugging Face," it is describing a release that never happened. Second, the model you actually want to download is Wan 2.2 — and it's genuinely good, genuinely free, and runs on surprisingly modest hardware.

---

Where the "Wan 2.7" Claims Come From {#where-27-comes-from}

There's a repeatable SEO pattern behind fake model versions, and once you've seen it you can't unsee it:

1. Register a domain around a plausible future version number. Wan ships fast, so "2.7" sounds believable.
2. Publish official-looking pages — a "docs" section, an "installation guide," a "download" button, sometimes a fabricated Apache 2.0 license claim.
3. Rank for the version queries ("wan 2.7 open source," "wan 2.7 weights") before the real vendor has anything to say about them.
4. Monetize the confusion — affiliate links, email capture, or just ad impressions from people looking for a model that doesn't exist.

We're not naming the specific domains — that's free promotion — but the same domain family runs this exact play on other model names too. We documented the same squatter pattern around Kimi K3, where speculative "guides" appeared for capabilities nobody had verified.

A reliable tell: check whether the page links to the vendor's actual Hugging Face or GitHub org — a real launch always does. The squatter pages rarely can, because the org would contradict them.

---

What Wan Has Actually Released {#official-releases}

The ground truth lives in exactly two places: the official Wan-AI org on Hugging Face and the official Wan-Video org on GitHub (Alibaba Cloud's video-generation org). As of July 20, 2026, here's what they contain:

• Wan 2.2 — the current open-weight flagship line, in three variants:
  • T2V-14B — text-to-video, 14B parameters
  • I2V-14B — image-to-video, 14B parameters
  • TI2V-5B — a smaller combined text/image-to-video model
• Wan 2.1 — the previous open generation, still maintained on the org
• Wan-Dancer-14B — the July 13, 2026 music-to-dance release (more below)

The main GitHub repos carry the Apache-2.0 license — so yes, the real open Wan models are permissively licensed. The fabricated part of the "Wan 2.7, Apache 2.0" claim isn't the license style; it's the version number.

What about 2.5 and 2.6? Those versions reportedly shipped as API/commercial-first releases — you could use them through hosted services, but no open weights were published. That's worth internalizing, because it makes the "2.7 open source" story even less plausible: the vendor's recent pattern has been to hold newer versions back from open release, not to quietly drop them on SEO blogs.

Honest aging clause: this is accurate as of July 20, 2026. If Wan does open-source a later version, it will appear on the official org linked above before anywhere else. Check there first — that habit alone makes you immune to the fake-version pattern.

---

What You Can Run Today, by VRAM {#run-by-vram}

Here's the practical part. The numbers below are community-reported — consistent across user reports, but not official specs, so treat them as a starting point rather than a guarantee.

A few pointers from that table:

• If you have 16GB of VRAM, the 14B at Q5_K_M is the quality play — FP8 and GGUF quantizations are what make the 14B models practical on consumer hardware at all.
• If you have 8GB, don't fight the 14B — the TI2V-5B exists precisely for you, and it handles both text-to-video and image-to-video.
• Below 8GB, offloading works but tests your patience. It's a "prove it runs" setup, not a workflow.

For the full installation walkthrough — ComfyUI setup, workflows, and prompt technique — see our Wan video generation guide. If you're still choosing between model families, our local AI video generation overview compares the whole field, and the HunyuanVideo guide covers the other heavyweight open video model worth your VRAM.

---

Wan-Dancer-14B: The Real New Release {#wan-dancer}

While the internet argued about a fictional 2.7, the Wan team shipped something real: Wan-Dancer-14B, released July 13, 2026. It's a music-to-dance video model — feed it audio, get dance video out — and it appears on the official GitHub org, which is more than "Wan 2.7" can say.

Before you get excited about running it locally, one number from the release materials: the reference rig is 8x A800 GPUs. That's datacenter-class hardware. Per the release materials, this is a model built to demonstrate capability, not to slot into a ComfyUI workflow on a 4070.

Will community quantizations eventually shrink it to consumer cards, the way GGUF did for the 2.2 line? Maybe — but as of July 20, 2026, any claim that Wan-Dancer runs on desktop hardware is unverified. Apply the same filter you'd apply to the 2.7 pages: if it's not demonstrated on the official org or reproducible community builds, it's marketing.

---

How to Verify Any "New Model" Claim {#verify-releases}

This keeps happening — fake Stable Diffusion versions, fake Wan versions, speculative Kimi guides — so here's the 60-second check that settles it every time:

1. Go to the vendor's own Hugging Face org. For Wan, that's huggingface.co/Wan-AI. Weights either exist there or they don't.
2. Check the vendor's own GitHub org. For Wan, github.com/Wan-Video. A real release has a repo, a license file, and commit history.
3. Distrust any page that claims a release but links to neither. That's the signature of the squatter pattern, not of a launch.

Aggregator sites, "docs" mirrors, and version-number domains are downstream at best and fabricated at worst. The org pages are the source of truth — everything else is commentary.

---

Verdict {#verdict}

Wan 2.7 is not open source, because Wan 2.7 does not exist. As of July 20, 2026, the official open-weight line stops at Wan 2.2, versions 2.5 and 2.6 reportedly stayed API-only, and the "2.7" pages ranking for your search are SEO fabrications following a known squatter playbook.

The good news is that the real answer is better than the fake one. Wan 2.2 is Apache-2.0 licensed on the official repos, ships in three variants, and — per community reports — scales from a 16GB card running the 14B at 720p down to an 8GB card running the TI2V-5B. That's a genuinely strong local video stack, available today, from the official org.

Skip the version-number domains. Download from the official org, start with our Wan setup guide, and if a real Wan 2.3-or-later open release ever lands, you'll see it on the org before any SEO site tells you about it.

---

Sources {#sources}

• Wan-AI on Hugging Face — official model weights (stops at Wan 2.2 as of July 20, 2026)
• Wan-Video on GitHub — official Alibaba Cloud org: Wan2.1, Wan2.2, Wan-Dancer repos, Apache-2.0

---

FAQ {#faq}