How to Run Wan 3.0 on QwenCloud: Model IDs, Pricing and Async API Setup (2026)

Most people who ask how to run Wan 3.0 from QwenCloud are stuck on something smaller than they think. They have an account, they have a prompt, and they are staring at a 404 or an empty response because the model ID they guessed does not exist, or because they called a synchronous endpoint on a job that takes minutes, or because they picked the tier that costs twice as much without noticing.

I went through it end to end: the model catalog, the real per-second prices (including the 30%-off promo that quietly applies to one tier and not the other), the async call shape, the rate limits that decide whether your batch runs at night or next week, and the cost arithmetic for a 30-second clip before you spend money finding out.

This is the setup guide, with the numbers filled in. Prices below were captured from QwenCloud's model pages in early November 2026 — check them again before you build a budget, because Wan pricing has moved twice this year.

TL;DR

• There are three Wan model IDs on QwenCloud, and picking the wrong one doubles your bill. wan3.0-video is the standard all-in-one model; wan3.0-video-prime is the high-speed variant with aligned capabilities at roughly 2x the price; wan2.7-t2v is the previous generation at an in-between rate.
• wan3.0-video is currently 30% off: $0.05 → $0.035/sec at 480P, $0.10 → $0.07/sec at 720P, $0.20 → $0.14/sec at 1080P. The Prime tier lists $0.068 / $0.14 / $0.28 per second with no discount showing.
• Wan 3.0 generates up to 30 seconds in one call, which is the actual reason to move off Wan 2.7 — the 15-second wall was the biggest constraint on the previous generation.
• The endpoint is asynchronous. You must send X-DashScope-Async: enable, keep the returned task ID, and poll. A synchronous POST looks like a hang.
• Rate limits are per model and identical across the three: 5 concurrent tasks, 500 queued, 300 requests per minute. Five concurrent tasks is the number that decides how long a batch of 200 clips takes.
• For a 30-second 1080P clip, standard Wan 3.0 costs about $4.20 on promo versus $8.40 on Prime and $4.50 for two 15-second Wan 2.7 renders — the 30-second pass is not just cheaper, it removes a seam.

What Wan 3.0 Actually Is on QwenCloud

QwenCloud lists two Wan 3.0 entries plus the older Wan 2.7 text-to-video model. They are not aliases; they are separate deployments with separate prices.

Model ID	What it is	Inputs	Max length	Best for
wan3.0-video	All-in-one video model: reference, editing, replication and driving	Audio, image, text, video	30s	Most work — best price-to-capability ratio
wan3.0-video-prime	High-speed variant, capabilities aligned to the standard model	Audio, image, text, video	30s	Iteration speed when time costs more than credits
wan2.7-t2v	Previous generation, text-to-video only	Audio, text	15s	Cheap T2V when you do not need 30s or references

The interesting part is what "all-in-one" means here. The standard model's own description says it unifies reference, editing, replication and driving, accepts omni-modal references, and can parse files, web pages and complex images as input. In practice that means one endpoint covers the jobs that used to need four: an image reference for identity, a video reference for motion and style, audio for lip-sync, and an edit instruction for a change inside an existing clip.

Prime is the same capability set with faster end-to-end generation. The docs describe it as "capabilities aligned to the standard version", which is a polite way of saying: same model, more compute, double the price. That is a real trade if you are iterating on a prompt twenty times, and a waste if you are rendering a batch once.

The Pricing Table You Actually Need

Here is the complete published rate card, with the promo discount applied and a per-clip cost for the three durations people actually render.

Model	480P	720P	1080P
wan3.0-video (list)	$0.05/s	$0.10/s	$0.20/s
wan3.0-video (30% off)	$0.035/s	$0.07/s	$0.14/s
wan3.0-video-prime	$0.068/s	$0.14/s	$0.28/s
wan2.7-t2v	—	$0.10/s	$0.15/s

And the same numbers converted into something you can put in a budget:

Job	wan3.0-video (promo)	wan3.0-video-prime	wan2.7-t2v
5s · 720P	$0.35	$0.70	$0.50
15s · 720P	$1.05	$2.10	$1.50
30s · 720P	$2.10	$4.20	n/a (15s cap)
5s · 1080P	$0.70	$1.40	$0.75
15s · 1080P	$2.10	$4.20	$2.25
30s · 1080P	$4.20	$8.40	n/a (15s cap)

Three observations that change how you plan:

The promo is doing real work. At 30% off, 1080P Wan 3.0 costs $0.14/sec — less than Wan 2.7's $0.15/sec — for a newer model with double the maximum length. If the discount is live, there is almost no reason to run Wan 2.7 text-to-video. If it lapses, the two are nearly identical at 1080P and 3.0 still wins on duration.

Prime is only worth it when you are exploring. $8.40 for a 30-second 1080P clip is real money; you pay it to cut iteration latency, not to get better output. Draft on Prime for five short attempts, then render finals on standard.

480P is a legitimate drafting tier. At $0.035/sec, a 5-second 480P preview costs under four cents. Blocking out composition at 480P and re-rendering the keeper at 1080P turns a $4.20 habit into a $0.90 one.

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

Step 1 — Get an API Key and Pick a Region

Wan 3.0 runs on QwenCloud's Singapore deployment, ap-southeast-1. Keys are issued from the QwenCloud console (home.qwencloud.com/api-keys) and used as a bearer token.

export DASHSCOPE_API_KEY="sk-..."

Two things worth knowing before you start:

• The base URL is regional. The endpoint used in QwenCloud's own example is https://ap-southeast-1.maas.qwencloudapi.com/.... If you are used to DashScope's Beijing URL, that is a different service and a different key.
• Billing is per second of output, not per request, and it applies to failed-but-generated output the same way. Your first cost surprise will be a batch of rejects, not a big clip.

Step 2 — Make the Async Call Correctly

Video generation does not return in an HTTP response. The API expects the X-DashScope-Async: enable header, returns a task ID, and you poll for the result.

Here is the documented request shape, filled in with the 720P tier:

curl --location 'https://ap-southeast-1.maas.qwencloudapi.com/api/v1/services/aigc/video-generation/video-synthesis' \
  -H 'X-DashScope-Async: enable' \
  -H "Authorization: Bearer $DASHSCOPE_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "wan3.0-video",
    "input": {
      "prompt": "A slow dolly-in on a rain-soaked Tokyo alley at night, neon reflections on wet asphalt, a courier steps out of frame left. Cinematic, 35mm, shallow depth of field."
    },
    "parameters": {
      "resolution": "720P",
      "ratio": "16:9",
      "duration": 15
    }
  }'

The three parameters that matter:

Parameter	What it does	What to set
model	Picks the tier and the price	wan3.0-video unless you need speed
resolution	480P, 720P, 1080P	Start at 720P; use 480P for drafts
duration	Clip length in seconds	Up to 30 on 3.0, 15 on 2.7
ratio	Aspect ratio	16:9, 9:16, 1:1, or adaptive to inherit the input

Then poll the task endpoint until status flips to succeeded, and download the video URL immediately. Provider-hosted output URLs are not permanent; copy the file to your own bucket the moment the job completes.

Ready to try it yourself? Try Wan 3.0 Free →

A gotcha that costs people an afternoon: if you omit X-DashScope-Async, some clients get a fast error rather than a queue — which reads like "model not found" in a retry loop. If the response is instant and wrong, check the header first.

Step 3 — Plan Around Five Concurrent Tasks

Wan 3.0's published limits are the same on all three model IDs:

Limit	Value	What it means in practice
Concurrency	5	Five tasks running at once, per model
Async queue limit	500	You can enqueue 500 tasks before rejections
RPM	300	Request rate is almost never your bottleneck

Five concurrent tasks is the constraint that decides your throughput, and it is stricter than most people expect. A 30-second 1080P render might occupy one slot for several minutes. If a render takes three minutes, five slots give you roughly 100 clips an hour at best — and that is before any render runs long.

Practical consequences:

• Batch jobs need a scheduler, not a Promise.all. Queue tasks, run five, backfill as each completes.
• Per-model limits are separate, so wan3.0-video and wan3.0-video-prime can run five each. Prime is a legitimate way to buy extra throughput on a deadline, not just speed per job.
• Build a status table, not an in-memory map. With hundreds of queued tasks, a restart that loses state means re-billed renders.

Step 4 — Get the 30 Seconds You Paid For

The reason to be on 3.0 instead of 2.7 is the length ceiling. Two approaches, and the difference is bigger than it looks.

Two 15-second Wan 2.7 renders: $2.25 + $2.25 = $4.50 at 1080P, two seams to hide, two prompts to keep consistent, and the second clip's opening frame has to match the first clip's ending. In practice you will do three or four attempts on the second half.

One 30-second Wan 3.0 render: $4.20 at 1080P on promo, one call, no seam, and the model holds the through-line itself because it generated the whole thing.

That is the argument. The 30-second pass is cheaper than the two-hop equivalent and materially better, which is rare enough in this market to be worth stating plainly. If your content is 30-second social, ads or short narrative, moving to 3.0 is not a quality upgrade — it is a cost and complexity reduction.

For prompts at this length, the structure that works is time-coded, the way Alibaba's own examples are written: Shot 1 [0–3s]: … Shot 2 [3–6s]: … through to the end. A single paragraph prompt over 30 seconds tends to drift; shot-by-shot timing keeps it on rails. For a closer look at how it stacks up against other models, see Wan 3.0 vs Flux 3 Video. If you want to test it without installing anything, the free image-to-video generator works in the browser.

Step 5 — Decide Whether You Need QwenCloud at All

QwenCloud is the direct route, and it is not the only one. Where it wins and loses:

Route	Best for	Trade-off
QwenCloud API	Volume, automation, custom UI	You build the interface; per-second billing is unforgiving of retries
Hosted generator with credits	Creators who do not want an API key	Less control over exact parameters
Affiliate model sites	Trying Wan 3.0 before committing	Front-ends vary; check which tier is actually behind the button
Local weights	Unfiltered or offline work	Separate question entirely — see the notes below

On that last row: if what you actually want is generated content without provider-side moderation, that is not a setting on QwenCloud. Hosted Wan endpoints apply provider content policy, and the model itself has limits. Whether Wan 3.0's weights are downloadable at all is a separate question from whether the hosted API will render a specific prompt — our Wan 3.0 complete guide covers the model itself, and it is worth reading before you plan a workflow around a filter that does not exist.

Five Mistakes That Cost Real Money

1. Using Prime for batch renders. Double the price for the same output. Prime is for iteration, standard is for delivery.
2. Rendering at 1080P while you are still changing the prompt. Draft at 480P or 720P, promote the keeper. This alone cuts a project's spend by half.
3. Forgetting the async header. Instant errors get misread as model availability problems.
4. Holding output URLs instead of files. Provider URLs expire. Download on completion.
5. Treating five concurrency as five users. It is five tasks, model-wide, for your account. A single impatient user clicking generate eight times can starve your entire queue.

Try Wan 3.0 Without an API Key

If you want the 30-second ceiling and the reference inputs without writing a polling loop, the hosted route gets you there in a browser:

Run Wan 3.0 on the Free Generator

Everything the API charges you for, without the API. The free Wan video generator runs the Wan family in a browser, so you can test whether the 30-second output is worth the per-second price before you commit a billing account to it.

• Text-to-video and image-to-video in one place, including the reference inputs that make Wan 3.0 more than a prompt box.
• No API key, no region setting, no async polling — the three things that stop most creators from reaching the model at all.
• No concurrency queue of your own to manage; five-task limits are our problem, not yours.
• Credits instead of per-second invoices, which makes a wasted render cheap instead of expensive.
• A fast way to benchmark tiers: render the same prompt on 2.7 and 3.0 and compare, which is the only way to know whether you need the newer model.

If you are buying access rather than building it, Wan 3.0 through Pollo is the hosted option I keep coming back to for longer clips, and Wan 2.7 image-to-video is still the cheaper route for 15-second work.

The Bottom Line

Running Wan 3.0 from QwenCloud comes down to four decisions: which model ID (standard, unless you are iterating), which resolution (draft at 480P or 720P, deliver at 1080P), whether you need 30 seconds (if yes, 3.0 is cheaper than stitching two 2.7 clips), and how you handle the five-task concurrency ceiling.

Get those right and the actual bill for a 30-second 1080P clip is about $4.20 on the current promo — versus $8.40 on Prime and $4.50 for the two-hop Wan 2.7 equivalent. Get them wrong and the same job costs you three or four times that in retries before you notice.

Start small: one key, one prompt, one 480P draft. The setup takes twenty minutes; the pricing mistakes take weeks to find.

FAQ

How much does Wan 3.0 cost per second on QwenCloud? Standard wan3.0-video lists $0.05/sec at 480P, $0.10/sec at 720P and $0.20/sec at 1080P, with a 30% discount currently applied ($0.035 / $0.07 / $0.14). The high-speed wan3.0-video-prime tier lists $0.068 / $0.14 / $0.28 per second.

What is the difference between wan3.0-video and wan3.0-video-prime? Prime is the high-speed variant with capabilities aligned to the standard model — same 30-second ceiling and same input types, faster end-to-end generation, roughly double the per-second price. Use it while iterating, not for final batches.

Can Wan 3.0 really generate 30 seconds in one call? Yes. QwenCloud's model page describes both Wan 3.0 entries as generating videos up to 30 seconds, with the standard model supporting omni-modal reference input including files, web pages and complex images.

Why does my request fail instantly? The most common cause is a missing X-DashScope-Async: enable header — video synthesis is asynchronous, so a synchronous call returns an error rather than a task. The second most common cause is a key from the wrong region; the endpoint here is ap-southeast-1.

How many clips can I render at once? Five concurrent tasks per model, with 500 queued and 300 requests per minute. Prime and standard have separate limits, so running both gives you ten concurrent slots.

Is Wan 3.0 faster on Prime, or just cheaper? Faster, at about twice the price. Prime does not change output capability — same resolutions, same 30-second limit, same input modalities. If you are not blocked on latency, standard is the correct default.

Should I still use Wan 2.7? For 15-second text-to-video at 1080P, Wan 2.7 is $0.15/sec versus Wan 3.0's $0.14/sec on promo — so usually no. Wan 2.7 remains relevant for the specific modes it documents well, like continuation and audio-driven lip-sync, and for workflows already built around its API shape.

How do I get unrestricted output? You do not, on this endpoint. QwenCloud's hosted Wan models apply provider content policy, and there is no toggle that removes it. Whether model weights are available for local use is a separate question, covered in our Wan 3.0 guide.

Related guides

• Wan 3.0 complete guide — the model, the modes and what the 30-second pass changes
• Wan 3.0 vs Wan 3.0 Prime — when the tier upgrade actually pays
• Wan 2.2 vs Wan 3.0 — what two generations of difference look like in output
• Wan 2.6 free — credit math and free routes on the previous generation
• Free Wan 2.5 image-to-video — the cheapest image-to-video route if 3.0 is overkill

References

• QwenCloud — Wan3.0-Video — 30-second output, omni-modal inputs, 480P/720P/1080P pricing with 30% discount, rate limits, DashScope async call example
• QwenCloud — Wan3.0-Video-Prime — high-speed variant, aligned capabilities, $0.068/$0.14/$0.28 per-second pricing
• QwenCloud — Wan2.7-T2V — previous generation, $0.10/sec at 720P and $0.15/sec at 1080P, 15-second maximum
• Alibaba Cloud Model Studio — model overview — confirms the Wan3.0 / Wan3.0-Video-Prime / Wan2.7-T2V catalog and routing
• QwenCloud API keys and console — key issuance for the ap-southeast-1 deployment