🆚 Wan 3.0 vs Wan 2.7: What Actually Changed in 2026

Every generation bump in the Wan line gets the same treatment from the SEO farms: a "vs" page full of confident numbers that no Alibaba source has ever printed. We got burned checking one of those pages in April, so for Wan 3.0 vs Wan 2.7 we did it the slow way — official Alibaba Cloud blogs, Model Studio API docs, the Wan-AI Hugging Face org, and the Wan team's own X account. Everything below traces to one of those, and where it doesn't, we say so out loud.

The short verdict: Wan 3.0's beta is a genuine capability jump — one 30-second take and a document-to-video input mode nothing else in the line has — but Wan 2.7 is still the version with stable, fully documented production endpoints. Which one you should be on depends on whether you're shipping this week or building for next quarter.

By the numbers#

Every figure verified this session against the source in the right-hand column. Pricing hedged as of August 2026.

The baseline: what Wan 2.7 actually shipped#

Wan 2.7 landed in early April 2026 as two official announcements — Wan2.7-Image and Wan2.7-Video — and the Model Studio docs behind them are the most complete API surface the Wan line has ever had. The image-to-video reference lists three distinct tasks: first-frame-to-video, first-and-last-frame-to-video, and video continuation. There's a separate instruction-based video-editing endpoint that takes text, images and video for edits and style transfer. And 2.7 replaced the old size field with resolution tier plus ratio, which is a small change that matters a lot if you template your API calls.

That endpoint list is 2.7's real identity. It isn't the longest or flashiest generator — it's the one where a keyframe pipeline, a continuation chain and a repair pass all have documented, stable contracts. On our rig, that's what kept it in rotation even after Seedance 2.5 took over our default text-to-video slot.

Worth hearing a skeptical take before the 3.0 hype recalibrates your memory of 2.7 — this reviewer stress-tested it at launch and found the gap between demo reel and real output:

His conclusion matches our April notes: strong identity handling and editing, motion that occasionally goes uncanny on complex physics. Keep that in mind, because 3.0's pitch aims squarely at the motion side.

What Wan 3.0 actually adds#

Three things, per the official announcement and corroborating coverage — and only three we'd call confirmed.

One ~30-second take. The headline. Wan 2.6 topped out at 15 seconds; 2.7's docs lean on continuation to chain longer sequences. Wan 3.0 claims roughly 30 seconds in a single generation pass, which is a different creative tool entirely: continuous camera language, one-take shot grammar, no stitch seams to hide. If it holds coherence across the full duration — the thing every long-clip model fumbles — it's the biggest single-spec jump in the line's history. Our full accounting of what's confirmed versus claimed lives in Wan 3.0: what's actually confirmed.

Document-to-video. The beta accepts documents, spreadsheets and slide decks as inputs alongside text, image, audio and video. Nobody else ships this. For explainer and marketing work — feed it the deck, get the video — it's potentially the most commercially interesting thing in the release, and it has no equivalent in 2.7's multimodal input list.

"Reality-grade rendering." Alibaba's phrase, not a spec. We list it because it's official; we can't quantify it because no benchmark accompanies it.

The announcement itself, from the horse's mouth:

What 3.0 does not visibly add yet: the mature endpoint family. We haven't seen official first-and-last-frame or video-editing docs for wan3.0-video the way 2.7 has them. Betas grow surfaces over time, but today, if your pipeline depends on those contracts, 2.7 is the version that honors them. Access paths, region quirks and the beta signup flow are in how to access Wan 3.0.

The fabricated-spec problem, again#

Search "wan 3.0 vs wan 2.7" today and most of page one is a cluster of lookalike domains — wan27.org, wan30.net, wan3pro.video, wan30ai.com, morphic.com, youart.ai, anikuku.com — none of them affiliated with Alibaba. In one search session we watched these sites claim, with straight faces: Wan 3.0 renders native 4K (no Alibaba source prints this, and the beta pricing tiers third parties read from Model Studio stop at 1080p); Wan 3.0 shipped as Apache 2.0 open weights in 1.3B and 14B sizes (flatly false — see below); and a tidy "$12.00 per minute vs $9.00 per minute" pricing comparison built on numbers that appear on no official rate card we can find.

This is the same playbook that manufactured a fake "Wan 2.7 vs 2.6" spec sheet back in April, complete with a 30fps figure we've never seen an Alibaba or partner source confirm — partner listings for 2.6 say 24fps. These sites exist to catch version-number searches before official docs rank. Our rule, which we'd suggest adopting: if a Wan number doesn't appear on alibabacloud.com, huggingface.co/Wan-AI, github.com/Wan-Video, or the @Alibaba_Wan account, it doesn't exist yet.

For an early hands-on read on the beta that's actually watching output instead of inventing spec sheets:

The "Seedance killer" framing is doing a lot of work in that title — our own head-to-head on that question is in Wan 3.0 vs Seedance 2.5.

The open-weights reality check#

The Wan brand earned its goodwill by open-sourcing Wan 2.1 and 2.2, so people keep assuming each new version drops on Hugging Face. It hasn't happened since 2.2. We checked both huggingface.co/Wan-AI and github.com/Wan-Video on August 10, 2026: there are no weights for 2.5, 2.6, 2.7 or 3.0. The newest open release is Wan2.2-Animate-2-14B, updated around August 9 — the animation/motion-transfer line, not the flagship generator.

So the honest framing of this comparison is API-vs-API. If open weights are your requirement, this whole article is moot and your model is still Wan 2.2 — our local setup guide covers running it on consumer hardware, and the broader landscape is in open-weight AI video models. Anyone telling you Wan 3.0 is Apache 2.0 is reading a fabrication site.

Who should switch, who should wait#

Test the 3.0 beta now if:

• Your bottleneck is clip length. A ~30-second native take beats any continuation chain for motion coherence, in every long-clip model we've tested.
• You make explainers, product walkthroughs or slide-driven content. Document-to-video is a workflow shortcut nothing else offers, and being early on it is cheap.
• You're evaluating platforms for Q4 and want to know where the line is heading.

Stay on 2.7 for now if:

• Your pipeline pins first or last frames, chains continuations, or uses the video-editing endpoint. Those contracts are documented on 2.7 and not yet visible on 3.0.
• You bill clients against predictable per-clip costs. 2.7's pricing is established; 3.0's beta rates are provisional and could move at general availability.
• You can't absorb beta flakiness — queue times, quota changes, silent model updates — in a delivery schedule.

Skip both if you need weights on your own GPUs. That's Wan 2.2, full stop — and for a cloud alternative with native audio at 2K, our MiniMax H3 vs Wan 2.2 comparison is the relevant read.

Our pick#

For production this month: Wan 2.7. Boring answer, but endpoints beat headlines when an invoice is attached, and 2.7's first/last-frame and editing APIs are the parts of the Wan line we actually route jobs through. For where we're spending our own experiment budget: the 3.0 beta, without hesitation. A real 30-second single take plus document input isn't an increment, it's a new category of job the line can do — and if the beta's coherence holds up at general availability, this comparison flips. The full feature walkthrough, prompt behavior and our test clips are in the Wan 3.0 complete guide.

How we checked#

Wan 2.6 and 2.7 facts come from Alibaba Cloud's official blog posts (602742, 602995, 603009) and the Model Studio API reference pages for text-to-video, image-to-video and video editing, read August 10, 2026. Wan 3.0 claims come from the official beta announcement and @Alibaba_Wan's post, with beta pricing tiers attributed to third-party trackers reading Model Studio's release log — we have not seen an official English rate card, so treat those yuan figures as provisional as of August 2026. Open-weights status was checked directly on the Wan-AI Hugging Face org and Wan-Video GitHub org the same day. We deliberately cited nothing from the wan27.org/wan30.net cluster except to document its claims as unverified. We have run beta generations through third-party access, not yet through a direct Model Studio account; duration and input-mode claims are Alibaba's, corroborated by independent coverage, not our own measurements.

Last verified: August 10, 2026.

The number to watch isn't 30 seconds — the beta already printed that. It's what the 1080p tier costs when "beta" comes off the label.