Wan Video Models Compared: Wan 2.1 to Wan 3.0 - Which Should You Use in 2026?

Introduction

"Wan video" used to mean one thing. In 2026 it means at least six different products that share a logo, and choosing the wrong one is the most common reason people think the Wan family is overpriced, slow, or impossible to run.

I ran into this the hard way. I spent a week trying to download Wan 2.7 weights — a model I had read was open — before accepting the obvious: there are no Wan 2.7 video weights. The open Wan line stops at Wan 2.2. Everything since is served through someone's API, and every API prices, limits, and filters differently.

This guide is the comparison I wish I had before that week: what each Wan version actually is, what shipped as open weights versus hosted API, what each one can do (duration, resolution, audio, control), what it costs per second, and a simple decision framework for picking the right one for your job.

TL;DR

The open Wan line ends at Wan 2.2 (Apache 2.0). Wan 2.5's weights never shipped, Wan 2.6 was API-only, Wan 2.7's video models are hosted endpoints, and Wan 3.0 has no public checkpoints.
Duration is the headline ladder: Wan 2.1/2.2 do roughly 5-second clips, Wan 2.7 does 15 seconds, Wan 3.0 does 30 seconds in a single pass.
Audio is the second ladder: 2.1/2.2 are silent, 2.6 needs a separate audio workflow, 2.7 adds dialogue and lip-sync editing, and 3.0 generates sound in the same pass as the picture.
Cost now spans 10× depending on channel. Discount tiers on 2.5/2.6 land around $0.02–$0.07 per second, hosted 2.7 runs about $0.10/second (roughly $0.086/s at 720p, $0.144/s at 1080p on resellers), and Wan 3.0 lists at $0.20/second at 1080p — about $6.00 for a full 30-second clip.
Free is still on the table, but only for the older generation. Browser tools running Wan 2.1/2.2 cover short clips and image-to-video at no per-second cost — see the free Wan video generator.
Version claims need verification. Before believing any "Wan X is open" post, check the official Hugging Face and GitHub organisations. That check has been wrong-proof for two years running.

The Whole Wan Family in One Table

Version	Released	Max clip	Resolution	Audio	Weights	Where to run it
Wan 2.1	Feb 2025	~5s	480p / 720p	No	Open — Apache 2.0	Local (ComfyUI), free browser tools, self-host
Wan 2.2	Jul 2025	~5s	480p / 720p	No	Open — Apache 2.0	Local, incl. 5B variant that runs on a single 24 GB GPU
Wan 2.5	Sep 2025	Short-form	720p class	Separate workflow	Announced as open, never shipped	Commercial API only
Wan 2.6	Dec 2025	Multi-shot storytelling with shorter clips	720p / 1080p	Separate workflow	Closed	Alibaba Cloud Model Studio and partners
Wan 2.7	Mar–Apr 2026	15s (t2v/i2v 2–15s; reference-to-video and editing 2–10s)	720p / 1080p @ 30fps	Dialogue and lip-sync editing, driving-audio input	Closed (video); Wan 2.7-Image released alongside	Hosted APIs, consumer platforms
Wan 3.0	Beta 6 Aug 2026	30s single pass	480p / 720p / 1080p (no 4K tier)	Generated in the same pass	Closed	Alibaba Cloud Model Studio, Qwen Cloud, partners
Side branches	Jul 2026	Varies	Varies	Varies	Open — Apache 2.0	Local + hosted: WanSong (music), Wan-Dancer-14B (dance), Wan-Streamer (real-time streaming), Wan2.2-Animate-2 (character animation)

Read the "Weights" column before anything else, because it is the line that determines whether your cost is a subscription or an electricity bill.

Quick Verdict: Pick by Job, Not by Version Number

Newer is not automatically better for your use case. This is the table I actually use.

Your job	Use	Why
Test 30 prompt ideas this week on a budget	Wan 2.1/2.2 free browser tools	Zero per-second cost; good enough to find the shot
Product still that needs motion	Wan 2.7 image-to-video	First-frame and first-and-last-frame control, 1080p
15-second ad beat with spoken lines	Wan 2.7	Native dialogue and lip-sync editing without a second audio pass
One-take 30-second spot with sound	Wan 3.0	30 seconds and audio arrive in a single generation
Deck, spreadsheet or PDF to video	Wan 3.0	Document-to-video is unique to it
Unlimited generation with no per-clip billing, no filters	Wan 2.2 local (ComfyUI)	Apache 2.0, runs on your own GPU
Music video / dance choreography	WanSong, Wan-Dancer-14B	Purpose-built branch models, open weights, free locally
Character animation from a driving video	Wan 2.2-Animate-2-14B	Open checkpoint, released August 2026
Real-time / streaming video generation	Wan-Streamer	Built for continuous generation

The pattern: use hosted Wan 2.7 and 3.0 for the shots that need length, audio, or reference control, and use free or local Wan 2.1/2.2 for everything you are still testing.

If you want output today, start here: Launch Wan 3.0 Now →

The Three-Tier Map That Explains Every Pricing Argument

Almost every contradictory claim about Wan online comes from mixing three access tiers that behave differently:

Open weights (Apache 2.0) — Wan 2.1, Wan 2.2, and the 2026 branch models. You download them, run them locally, and pay nothing per generation. No platform filter, no rate limit, no queue. The trade is capability: these are the 5-second, silent generation models, and quality sits a generation or two behind the hosted flagships. This is what built the ComfyUI ecosystem around Wan.
API-only — Wan 2.5 (preview), 2.6 and 2.7. Priced per second or per token through Alibaba Cloud Model Studio and resellers. This is the production tier: documented endpoints, predictable output, per-second billing. Wan 2.7's surface is the most complete the Wan line has had — text-to-video, image-to-video with first-frame or first-and-last-frame input, video continuation, and instruction-based video editing.
Gated beta — Wan 3.0. Launched into public beta on 6 August 2026, hosted by Alibaba and partner platforms. It doubles the single-pass length to 30 seconds, generates audio in the same run, accepts document inputs, and holds up to 20 mixed references. It also has no open weights, which means "downloading Wan 3.0" is not a real option.

Two practical consequences:

If a page says "Wan 3.0 free download," it is wrong — the check takes thirty seconds: look for a wan3 repository in the official Hugging Face or GitHub organisations. There isn't one. Anything sold as Wan 3.0 weights has provenance you cannot verify.
If a page says "Wan can't do audio," it is describing tier one. Audio arrived with the hosted tiers, and tier three generates it natively.

How I Verify a Wan Version Claim (30-Second Check)

I do this before writing anything about a Wan release, because the misinformation volume is genuinely high:

Check the official model organisations. Hugging Face hosts the Wan-AI organisation and GitHub hosts the Wan-Video organisation. If the version number is not in either repository list, the weights do not exist — regardless of what a blog post says.
Check the API model list. Alibaba Cloud Model Studio's model list is the authoritative source for what is actually callable. A version that is not in that list is either unreleased or private.
Compare the dates. The most useful fact I have found in the last year: Wan 2.2's Animate-2 checkpoint shipped to Hugging Face one day after Wan 3.0's closed beta opened. Open releases and closed betas run in parallel, which is exactly why "latest" and "latest open" are different questions.

This is the difference between writing an article and copying one. It also saved me from paying for a "Wan 3.0 weights" bundle that turned out to be a re-cut Wan 2.2 checkpoint.

What Each Generation Actually Changed

Step	Motion and duration	Audio	Control	Practical upgrade
2.1 → 2.2	Same ~5s envelope, better motion stability	None	Basic prompt control	Smaller 5B variant runnable on one consumer GPU
2.2 → 2.5	Longer-form ambitions, open release promised	Separate workflow	API surface only	Mostly a commercial milestone, not a capability leap
2.5 → 2.6	Multi-shot storytelling with shorter clips	Separate workflow	Reference-to-video	Better narrative sequencing, still no native sound
2.6 → 2.7	15 seconds per clip	Dialogue and lip-sync editing, driving audio from your own file	Thinking mode, up to five consistent characters with individual voices, video continuation	The first version that removed the separate audio pass from most workflows
2.7 → 3.0	30 seconds per clip	Sound generated with the picture	Up to 20 mixed references (image, video, audio, document, webpage), interval-level regeneration including dialogue	One generation replaces a stitched sequence plus an audio pass

If you are on Wan 2.2, the two reasons to move to a hosted tier are length and audio. If you are on 2.7, the reason to consider 3.0 is reference capacity and document input, not the extra fifteen seconds on its own.

Cost: The Real Numbers Per Finished Clip

Published per-second rates, translated into what a single deliverable costs:

Route	Per-second rate	10-second clip	30-second clip
Free browser tools (Wan 2.1/2.2)	$0	$0	Not applicable (5s clips)
Local Wan 2.2 (own GPU)	Electricity + your time	$0 marginal	Not applicable (5s clips)
Discount tiers on Wan 2.5 / 2.6	~$0.02–$0.07/s	$0.20–$0.70	$0.60–$2.10
Wan 2.7, hosted reseller at 720p	~$0.086/s	$0.86	$2.58
Wan 2.7, hosted reseller at 1080p	~$0.144/s	$1.44	$4.32
Wan 2.7, flat aggregator rate	~$0.10/s	$1.00	$3.00
Wan 3.0 at 1080p list	$0.20/s	$2.00	$6.00
Wan 3.0 at 480p draft	~$0.05/s	$0.50	$1.50

Now scale it, because this is where the version choice stops being an aesthetic question:

Ten 15-second 720p clips a day on Wan 2.7 — 10 × 15 × $0.086 × 30 days = about $387 a month.
Ten 30-second 1080p clips a day on Wan 3.0 — 10 × 30 × $0.20 × 30 days = about $1,800 a month.
The same testing volume on free browser tools — $0, with a shorter duration ceiling.

Ready to try it yourself? Try Wan 3.0 Free →

The cost-correct workflow most teams land on: test on free Wan, draft on 480p where a draft tier exists, and pay for 1080p only on the generation that is actually shipping. Wan 3.0's 480p draft at roughly $1.50 for thirty seconds makes that strategy viable at the top tier too. For a closer look at how it stacks up against other models, see Wan 2.6 vs Wan 2.5 vs Wan 2.2.

Running Wan Locally: What Your GPU Can Handle

The local route only covers the open line, but the open line is more capable than people assume.

Hardware	Runs	Realistic output
RTX 4060-class, 8 GB	Wan 2.2 5B (quantised)	Short 480p/720p clips, slow
RTX 4070/4080-class, 12–16 GB	Wan 2.2 5B comfortably, 14B quantised with short clips	720p, a few seconds, workable iteration speed
RTX 3090/4090, 24 GB	Wan 2.2 14B quantised or full	720p at usable speed, the tier where local becomes practical
Multi-GPU	14B comfortably, plus branch models	Longer sequences, parallel batches

Two things the local route buys you that no hosted tier does: no per-generation cost, so unlimited iteration, and no platform content filter, since there is no platform. It also means no 30-second takes, no native audio, and setup time measured in hours rather than minutes.

If you want the free version of that experience without installing ComfyUI at all, browser tools cover the same generation envelope: start with the free image-to-video generator when you have a frame in mind, or the free text-to-video generator when you are starting from a prompt.

Common Mistakes with the Wan Family

Assuming the newest version is downloadable. Wan 2.7 and 3.0 have no public video weights. The last open release is the Wan 2.2 line, including Animate-2 from August 2026.
Confusing Wan 2.7-Image with Wan 2.7-Video. They are separate releases with separate access rules; image generation opened up more than video did.
Paying premium resolution during exploration. A 480p draft tells you whether the shot works. Re-render the winner.
Paying per second for shots the free tier already handles. Short social clips rarely need 30 seconds or native audio. Test free, ship paid.
Trusting a per-second price without checking the channel. A flat $0.10/second aggregator rate and a $0.144/second reseller rate are the same model at a 44% price difference.
Ignoring the branch models. If you need music, dance, or character animation specifically, WanSong, Wan-Dancer-14B, and Wan2.2-Animate-2 are open, free to run, and often the right tool rather than a general video model.

The Bottom Line

The Wan family is not one model with versions; it is three access tiers with very different economics. The open line — Wan 2.1, 2.2, and the 2026 branch models — costs nothing per generation and tops out at short, silent clips. The API-only line — 2.5 preview, 2.6, 2.7 — costs about $0.10/second and gives you 15-second takes, 1080p, and dialogue editing. The gated beta — Wan 3.0 — costs $0.20/second at 1080p and gives you a 30-second single take with sound generated in the same pass.

Want to see the difference on your own footage? Start creating with Wan 3.0 →

Pick accordingly: free Wan for anything you have not proven yet, Wan 2.7 for 15-second commercial work that needs control and speech, Wan 3.0 for one-take spots and document-to-video, and local Wan 2.2 when you need unlimited volume and full control more than you need the newest capability.

Start with Free Wan Video Generation

Try Wan Video Generation for Free

Test the shot before you pay per second for it. Wan 2.1 and 2.2 cover short clips, image-to-video, and still generation in the browser at no cost — the cheapest way to find out whether your prompt works at all.

No per-second billing while you iterate. Exhaust your ideas first, then spend on the version that fits the job.
Text-to-video, image-to-video, and still generation in the same place, so you can build the frame before you animate it.
Generate stills with the free Z-Image generator and turn them into motion, which is still the most reliable way to control composition.
No install and no account gymnastics — if you later want unlimited local generation, the same open weights that power these tools run in ComfyUI.
Upgrade only when the shot earns it — 15-second takes and native audio are a hosted-tier decision, not a starting point.

When you are ready for the hosted tiers, Wan 2.7 via a subscription host is the practical middle step before paying the 30-second premium.

Related guides

FAQ

Is Wan 3.0 open source?

No. Wan 3.0 entered public beta on 6 August 2026 and is served through Alibaba Cloud Model Studio, Qwen Cloud, and partner platforms. No wan3 checkpoints exist in the official Hugging Face or GitHub organisations — the newest open release is the Wan 2.2 line, including Wan 2.2-Animate-2-14B from August 2026. Any site selling "Wan 3.0 weights" is selling something unverifiable.

What is the latest version of Wan video?

Wan 3.0 is the newest version, in public beta since 6 August 2026, with a 30-second single-pass ceiling, native audio, up to 20 mixed reference inputs, and document-to-video. Wan 2.7 (March–April 2026) remains the most battle-tested production API with 15-second clips and dialogue editing.

Which Wan model is free?

Wan 2.1 and Wan 2.2 are free to run: they are Apache 2.0 open weights, so they cost nothing beyond hardware if you run them locally, and browser tools such as the free Wan video generator run them at no charge with no install. The hosted generations (2.5, 2.6, 2.7, 3.0) are all paid per second or per credit.

What is the difference between Wan 2.7 and Wan 3.0?

Two things: length and audio. Wan 2.7 tops out at 15 seconds with dialogue and lip-sync editing handled as a separate step, and supports up to five consistent characters. Wan 3.0 doubles the single-pass length to 30 seconds, generates sound in the same pass as the picture, accepts up to 20 mixed references, and adds document-to-video. Wan 3.0 costs about twice as much per second.

Can I run Wan video generation on my own GPU?

Yes — Wan 2.1 and Wan 2.2, plus the 2026 branch models, are Apache 2.0 open weights. An 8 GB card handles the quantised 5B variant, 12–16 GB is comfortable for 5B and workable for quantised 14B, and a 24 GB card (3090/4090) is where local Wan becomes practical at 720p. Wan 2.7 and 3.0 cannot be run locally because their weights were never released.

How much does a Wan video cost?

It depends entirely on the tier. Free for Wan 2.1/2.2 locally or in the browser; roughly $0.02–$0.07 per second on discount tiers for Wan 2.5/2.6; about $0.10/second for hosted Wan 2.7 (so roughly $1.50 for a 15-second clip); and $0.20/second at 1080p for Wan 3.0, which puts a full 30-second generation at about $6.00, or roughly $1.50 as a 480p draft.

Which Wan model should I use for image-to-video?

Free browser tools running Wan 2.1/2.2 for testing, and Wan 2.7 for production work that needs 1080p, first-and-last-frame control, or video continuation. Image-to-video is consistently more controllable than text-to-video because you supply the composition and the model only supplies motion.

References

Free Tools

Latest Posts

Start Creating

Ready to Create with Wan 3.0?

Try Wan 3.0 for AI video generation — start free in your browser, no setup required.

Text to Video

Image to Video

No Setup Required

Free to Try