Which Wan Can You Actually Use in 2026? Open Weights vs API vs the 3.0 Beta

Search for "Wan video model" in August 2026 and the results describe four different products as if they were one. There is an open-source Wan you can download, an API-only Wan that took over three versions ago, and since last week a gated beta that generates 30-second single takes. Articles routinely attribute one tier's capabilities to another, and at least one aggregator promised an open-source rollout of a version that never shipped weights at all.

Here is the actual map, from the official releases, the Model Studio catalog, and the reseller price sheets — with every number attributed to its channel, because in this family the same model has three prices depending on where you buy it. We operate Chinese video APIs in production (that muscle is documented in our MiniMax H3 manual), and Wan's operational quirks rhyme with what we run daily.

The three-tier map

Tier	Versions	Duration	Resolution	Audio	Access
Open weights (Apache 2.0)	2.1, 2.2	~5s fixed	480P/720P	Silent	HuggingFace download
API-only	2.5-preview, 2.6, 2.7	2–15s	720P/1080P, 30fps	Native + auto-dub	Alibaba Cloud Model Studio
Gated beta	3.0	up to 30s, one pass	480P/720P/1080P	Native	Apply for access

The fork happened in mid-2025: Wan 2.2 (July 2025, MoE models plus a 5B variant that runs on a single RTX 4090) was the last open release. The 2.5 preview arrived that September as API-only, and a year later it is still served under preview model IDs — followed by 2.6 (December 2025), 2.7 (April 2026), and now 3.0, none with weights. Four consecutive hosted-only releases is a direction, not an accident.

The open tier: free, silent, five seconds

If you own a GPU, Wan 2.1/2.2 is the floor for a zero-marginal-cost product-video pipeline: the 1.3B model needs 8.19GB of VRAM, and a 4090 renders a 5-second 480P clip in about four minutes. The catch people discover after downloading: open Wan is silent by default — native synchronized audio only exists from the API-only 2.5 preview onward. And the "unlimited length 1080P" line that circulates refers to the VAE encoder's capability, not generation: open models output about 5 seconds. Good for b-roll and background loops; not for a talking ad.

The API tier: where sellers actually land

The working models live in Alibaba Cloud Model Studio (international, Singapore endpoint) and the mainland Bailian platform. Operational facts that will save you a day each:

Wan 2.6 spans text-to-video, image-to-video, and R2V — upload a reference video of a person and it carries their appearance and voice into new scenes, which Alibaba bills as China's first reference-to-video model. For a seller that means a consistent presenter without re-filming. Two footnotes: the cheap wan2.6-i2v-flash tier generates audio by default (pass audio=false for silent), and per Gate.AI's reading of Alibaba pricing, R2V bills up to 5 seconds of your reference on top of the output — a "10-second" clip can meter as 15. Input video metering is becoming the industry norm; we flagged the same pattern on Seedance and H3 in our cost comparison.

Wan 2.7 (April 2026) adds multi-shot narrative, video continuation, and driving audio — your own MP3/WAV, 2 to 30 seconds, up to 15MB, truncated or silence-padded to fit. Caps: 2–15s for t2v/i2v, 2–10s for r2v and video editing, all 720P/1080P at 30fps. The trap: the official DashScope SDKs stop at 2.6 — 2.7 is raw HTTP with a new request protocol, so one product family contains two integration paths. Model IDs are dated snapshots (wan2.7-i2v-2026-04-25); pin them. Endpoints are region-locked (model, endpoint and key must match Singapore or Beijing), and result URLs reportedly live 24 hours — persist the MP4 on success.

Prices, by channel: OpenRouter serves 2.7 at a flat $0.10/second; resellers quoting the Alibaba route list $0.086/s at 720P and $0.144/s at 1080P (EvoLink, April 2026); discount tiers on 2.5/2.6 reach $0.02–0.07/s. Never write or budget "Wan costs X" without naming the channel — the spread is real and unresolved even between reseller price sheets.

Wan 3.0: the 30-second beta

Beta access opened August 6, 2026 (the launch event followed on the 10th — some outlets date the "release" from the event). The headline is legitimate: native 30-second single-pass generation, double the 2.7 cap, with continuous camera moves and no stitching. It also takes documents as input — a PDF, PPT or spreadsheet up to 100MB or 50 pages becomes a video brief — alongside text, images, video and audio.

List pricing: ¥0.30/¥0.60/¥1.20 per output second (480P/720P/1080P) on the mainland platform, about $0.05/$0.10/$0.20 internationally. A 30-second 1080P clip is roughly $6. Two hedges: access is gated behind an application, and beta model IDs and prices can change before GA. The only other production engine offering a 30-second single take today is Seedance 2.5, at a higher list rate — if your ad depends on one continuous shot, those are the two doors, and both are new.

Which Wan for product video

• Testing volume on your own GPU: open 2.2, silent clips, add voice in post.
• The seller workhorse: 2.6/2.7 API — product still in, 15 seconds out, native dubbing, your own voiceover via driving audio, a cloned presenter via R2V.
• The one-take ad: 3.0 beta, if your application clears.

The operational overhead — region-locked endpoints, SDK gaps, snapshot pinning, 24-hour URLs, reference metering — is the same class of work we absorb for the engines behind Riffkit: hand it a short video that already worked and it rebuilds that formula around your product, engine choice and per-second rate visible on pricing before you render. We publish maps like this because reading them is our day job either way.