Wan 3.0 Free: Thirty Seconds of Video, Once, in One Region — and a Reference Clip Spends It Too
Alibaba's Wan 3.0 landing page promises a 30-second free allowance; the API reference defines the meter as input plus output duration, so a 10-second reference clip leaves 20 seconds of output, and the whole grant is one 90-day allocation in Beijing only. Read on 31 October 2026.
- Model
- Wan 3.0
- Observed
The free allowance for Wan 3.0 is 30 seconds of video, and it is not 30 seconds of output. Alibaba's own API reference defines the billed duration as the input video plus the output video whenever a reference clip is attached, and caps the two together at 30 seconds. A ten-second reference clip therefore leaves twenty seconds of finished video, and when those thirty seconds are gone the grant does not refill: it is a single allocation, valid for 90 days, on one region out of the six the model is served from.

Everything below was read on 31 October 2026 from Alibaba Cloud's public pages as a guest, with no account and no API key. Two things were not reachable that way, and they are named where they matter: the Model Studio console's per-model quota panel, which sits behind a login, and the price a logged-in account is actually charged. Where a figure comes from a third party rather than from Alibaba, it is marked as such.
Thirty seconds, and the meter runs at both ends
The 30-second number is printed on Alibaba's visual-generation landing page, in the panel headed Wan 3.0: the tile reads "按量计费更高性价比 / 免费额度 30 秒即刻体验". That is the whole of what the marketing page says about it — a number and a unit, with no region, no expiry and no definition of what counts against it.
The definition is in the API reference, and it is narrower than the tile suggests. Three separate passages decide how much of the 30 seconds a request actually spends:
- The
durationparameter, for a request with no video input, accepts an integer from 2 to 30. With a video input, the rule changes: input video duration plus output video duration may not exceed 30 seconds. Passing-1hands the choice to the model. - The
usageobject, returned with every finished task, definesdurationas the total video duration in seconds — equal to the output duration when there is no reference video, and equal to input plus output when there is one. - A reference video may be up to five segments with a combined length of no more than 15 seconds, and a single segment must be between 1 and 15 seconds.
Two different rules are in play here and they happen to carry the same number, which is worth separating before the arithmetic starts. The 30-second limit in the duration parameter is a per-request ceiling. The 30 seconds on the landing page is the whole free grant. Because the ceiling equals the grant, a single well-formed request can spend everything an account was given — there is no shape of request that is too large for the free tier, only one that is exactly large enough to end it.
Put those together and the arithmetic is forced. A ten-second reference clip plus a twenty-second output is 30 seconds of metered duration, which is the entire free allowance in one call. Spend the full 15 seconds of reference video the API permits and only 15 seconds of new footage remains inside the grant. The same page also caps reference audio at five segments and 15 seconds combined, and reference images at ten, so a request that leans on the reference mode is the most expensive shape a free account can send.
The reference and frame modes cannot be mixed either. Passing reference_image, reference_video or reference_audio together with first_frame or last_frame fails with InvalidParameter and the message "The two modes are mutually exclusive. Do not pass reference_xx and first_frame/last_frame at the same time." That matters for anyone planning a free run: the cheapest way to stay under 30 seconds is a plain text prompt, and the moment a reference clip enters the request, half the allowance can be spent before a single frame is generated.
One region out of six
The API reference lists six endpoints for this model — Beijing, Singapore, Tokyo, Frankfurt, Virginia and Hong Kong — and the choice is not cosmetic, because a request must match the region of the model and the API key or it fails outright.
Only one of those six carries the free allowance. Alibaba's free-quota documentation states it plainly: "仅华北2(北京)地域模型享有免费额度,其他地域无免费额度" — only models in the Beijing region have a free quota; every other region has none. The same rule is repeated in the troubleshooting section, where a missing quota is explained as either expiry or a region outside Beijing.
This is the part the tutorial videos miss most often, and it is easy to see why. An account created against the Singapore endpoint looks identical in every other respect: same model name, same parameters, same 30-second ceiling on a single generation. It simply has no free grant to spend, so the first successful call is a billed call. A transcription of the model page published on Alibaba's developer community on 8 September 2026 lists Singapore, Tokyo, Frankfurt and Virginia as the regions without quota and does not mention Hong Kong, which the API reference does serve — the two lists are not the same length, and the shorter one is the one being quoted around.
Ninety days, one grant, then a 403
The allowance is not a daily or monthly refresh. The free-quota page gives the validity as 90 days, counted from whichever is later of activating Model Studio, the model's release, or the approval of an application — and adds that a second registration under the same verified identity does not produce a second grant.
What happens at the end is configurable, and the default is the expensive one. The "免费额度用完即停" switch, which stops service instead of billing, is off unless the account owner turns it on; with it on, a call after exhaustion returns HTTP 403 with the error code AllocationQuota.FreeTierOnly. With it off, calls keep succeeding and the account is billed from the first second past the grant. The documentation recommends leaving it off in production, which is the right advice for a production account and the wrong advice for anyone testing whether the free tier is real.
One further claim comes from the developer-community transcription rather than from Alibaba's own documentation: that failed generations are neither billed nor counted against the free allowance. It is plausible and it is what most vendors do, but this page did not find it stated on any official page, so it is recorded here as a third-party statement.
The faster version lost its free quota, and the two pages now disagree
The standard model is wan3.0-video; there is a faster variant, wan3.0-video-prime. The 8 September transcription says the two "共享同一档 30 秒免费额度" — share one 30-second allocation.
Alibaba's free-quota page now carries a collapsed panel headed "已下架免费额度的模型列表" — the list of models whose free quota has been withdrawn — and expanding it changes that answer. The list opens with a date, 2026-10-29, followed by model codes, and wan3.0-video-prime is the second entry. wan3.0-video — the standard model — does not appear anywhere on the page. The rule the page gives for the list is that once a model is withdrawn, "新用户不再获得该模型的免费额度;已发放给老用户的额度仍可继续使用,直至到期或耗尽" — new users no longer receive that model's free quota, while quota already granted to existing users runs until it expires or is used up.
Two official statements also sit awkwardly beside the sharing claim. The same page's notes say free quotas for different models, including different snapshot versions of one model, are independent and are not pooled or shared. So the September transcription's "shared allocation" is contradicted by the general rule, and the October withdrawal list removes one of the two models from the arrangement altogether.
This page is not going to settle which reading is right, because Alibaba has not published a reconciliation. What can be said without guessing is narrower and more useful: an account created today, in Beijing, on the standard model, can expect 30 seconds to spend, and the faster variant is on the list of models that no longer hand that grant to new users. Anyone planning to route a free test through wan3.0-video-prime should check the console's quota panel first, because the September article that promised the sharing is now the older of the two documents.
Two official pages, two prices, a factor of 1.5 apart
Pricing is where the public pages stop agreeing with each other. The visual-generation landing page prints three per-second rates for Wan 3.0, by output resolution: ¥0.45 at 480P, ¥0.9 at 720P and ¥1.8 at 1080P, with no discount label on any of them.
The developer-community transcription prints a different set: ¥0.3, ¥0.6 and ¥1.2 per second, headed "原价" (list price), with a second column headed "折后价(按7折估算)" — discounted price, estimated at 70% — giving roughly ¥0.21, ¥0.42 and ¥0.84. Every pair differs by exactly a factor of 1.5.
One arithmetic coincidence is worth stating and then setting aside: ¥0.45 multiplied by 0.7 is ¥0.315, which truncated to one decimal place is ¥0.3. That would make the landing page's figure the list price and the transcription's figure the discounted one, with the transcription then applying the 70% a second time to produce its last column. It is a tidy explanation and it is only an explanation — neither page states it, and this page does not adopt it. Both sets of rates are recorded with their dates and their sources, and the difference stays visible.
What the rates do settle is the size of the free allowance in money. At the landing page's figures, 30 seconds of output is worth ¥13.50 at 480P, ¥27 at 720P and ¥54 at 1080P. That is this page's arithmetic on published rates, not a discount anyone promised, and it is the clearest way to say what "30 seconds free" is actually worth: about the price of one short clip, once.
There are no 3.0 weights to download
The open-weights route that made earlier Wan releases interesting does not exist for 3.0 yet. The official Wan-Video organisation on GitHub holds six public repositories; the newest by push date is Wan2.2, last updated 21 September 2026 under Apache-2.0. The other five are Wan2.1, Wan-Dancer, Wan-Animate-2, Wan-skills and a diffusers fork, none of them newer than Wan2.2. There is no repository for 3.0, and nothing in the organisation's listing suggests one is staged.
That is an observation about a listing on one day, not a statement about what Alibaba will or will not release. But it does close one of the search routes: anyone arriving at this query hoping to run Wan 3.0 locally on their own hardware has, as of today, nothing to download, and the 30-second allowance above is the only free path that exists.
What the "free unlimited" videos are actually measuring
A search for this model returns tutorials rather than documentation. Of the twenty video titles returned on 31 October 2026, nineteen contain the word "free", and several pair it with "unlimited". The comments underneath are less confident than the titles. Under one video titled "WAN 3.0 FREE UNLIMITED", a commenter writes "Fake, no unlimited, scam"; another on the same video writes "It's paid not free"; a third, in Roman Urdu, says the unlimited use requires a purchase first. Under a different tutorial, one comment reads "only 40 credits it gives to generate", and under another, "Why you keep saying 'for free' lol it costs 400".
None of those comments is about Alibaba's allowance. They are about credits on third-party sites that resell access to the model, where "unlimited" means a subscription tier and 40 credits is that site's own unit. The disagreement in the comments is real and the number in the comments is not transferable: a credit count on a reseller's page says nothing about the 30 seconds Alibaba grants, and no video in that result set quotes the 30-second figure, the Beijing-only rule, or the 90-day expiry. The gap is not that the videos are lying about a number; it is that they never reach the document where the number is defined.
What this page could not confirm
Three things were out of reach without an account, and they are listed rather than papered over.
The console's per-model quota panel was not read. The 30-second figure for the standard model therefore rests on the landing page's tile plus the 8 September transcription, not on a screenshot of the panel itself — and the API reference, read in full, contains no free-quota section at all. The API reference's silence is itself the finding: the document that defines the meter does not mention the grant.
Whether an existing account that already holds wan3.0-video-prime quota keeps it after the 29 October listing was not verified, and the documentation's own wording — existing grants run until expiry — is the only guide available.
The date 2026-10-29 at the head of the withdrawn list is reproduced as printed. Whether it marks the day those quotas were withdrawn, or the day the list was compiled, is not stated on the page, and the list itself runs back through models far older than that date.
Everything dated on this page was read on 31 October 2026. The full set of observations behind it, with channel and date on every row, is in the dataset.