Wan AI: Alibaba's video models, from open weights to Wan 2.7

The open-weight releases changed who gets to generate video. Wan 2.2 on a consumer GPU produces good image-to-video: subject consistency, real motion, zero per-clip cost, no queue, no content gatekeeping, full pipeline control in ComfyUI. Nothing commercial matches that combination for tinkerers and volume workflows.

The hosted line is quietly competitive: Wan 2.5 stepped up detail and motion stability, and Wan 2.7 added native audio and much longer clips (the two features that otherwise force you onto Veo or Sora) at aggressive API pricing.

Because the family spans free-local to premium-hosted, you can prototype locally and move the same prompting style to the hosted tiers when a project needs audio, length, or peak quality. Our Wan 2.2, 2.5, and 2.7 guides cover each rung in depth.

The setting that moved the needle for us on hosted generation: rendering at a lower base resolution, then running a 2x model-based upscale (we use RealESRGAN, not a simple resize) instead of generating at full resolution from the start. On a paid-per-second GPU, that cut our generation time roughly in half for a comparable final resolution, since the diffusion steps do far less work per frame and the upscale pass is cheap by comparison.

Local Wan has a hardware price of admission. A modern GPU with substantial VRAM, plus the willingness to set up ComfyUI. Cheapest per clip. Most expensive in setup effort.

Ecosystem polish trails the Western flagships. The consumer-facing apps and editing tools around Veo or Runway don't exist in the same form; Wan is a model you access through platforms and APIs more than a packaged creative suite.

Ceiling-level photorealism and prompt adherence at the very top end still belong to Veo 3 and the newest closed models; Wan's pitch is 90% of the result at a fraction of the cost, not outright leadership.

Local route: download the open Wan weights and run them in ComfyUI. Budget an afternoon for setup; after that every generation is free. Our Wan 2.2 guide is the walkthrough.

Hosted route: Alibaba's model API exposes the full line including 2.5 and 2.7 with per-second pricing, and a growing set of third-party generation platforms offer Wan models with credit systems. Often the easiest way to try 2.7's audio without an API account.

Against Veo 3: Veo wins on peak fidelity, ecosystem, and audio polish; Wan 2.7 answers with native audio and longer clips at a much lower price, plus the open escape hatch no Google product offers.

Against Kling and Hailuo: those have friendlier consumer apps and free daily credits; Wan counters with open weights and better economics at API volume. For a creator who wants to own their pipeline, Wan is the default choice.

In July 2026 the Wan team published Wan-Streamer, which is a different kind of thing from the clip generators above. It is a live video conversation: one model handling text, audio and video understanding and generation together, so you talk to a generated character and it answers in real time, seeing you through your camera while it does. In the demo it describes the person it is looking at unprompted, down to their shirt colour and the wall behind them, and holds normal back-and-forth. It is not tied to a fixed avatar: the character can be any person, a pet, or an anime figure described in words.

The numbers are the interesting part. v0.2 runs 640x368 at 25fps, up from 192x336, while holding roughly 200ms of model-side latency and about 550ms end to end. It gets there by splitting the work: a light "Thinker" handles perception, language and state at low latency, while a heavier multi-GPU "Performer" generates the video latents.

Version 0.3 followed on 16 July 2026 and kept those numbers identical, which tells you where the work went. Instead of chasing resolution, it reframes what a video is: a persistent "world" (the character's appearance, the room, the visual style) plus an "event stream" of things that change (speech, body movement, camera motion). During a live session your own camera, microphone and typed text join that event stream. The practical result is that the character stops being a talking head. Earlier versions could only do portrait shots of someone sitting and speaking; in v0.3 you direct the body by voice mid-conversation. Ask it to make a peace sign, put its hands on its hips, fix its hair or look down at the floor, and it does, while still holding the conversation.

We build real-time companion products, so here is the honest read on those numbers. Around 550ms round-trip is conversational; it sits near the pause a person leaves before answering, and it is far better than the multi-second lag that makes video chat feel broken. The resolution is the part that is not there yet. 640x368 is a small window, fine for a phone-sized talking head but short of the mid-shot framing the demo is reaching for. Voice-directed body control is the more meaningful upgrade, because a companion that only ever sits still reads as a video call with a recording.

The catch that decides whether you can use any of this: there is only a technical paper, and that is now true three times over. v0.1, v0.2 and v0.3 arrived across roughly six weeks with no weights, no product and no API between them. The Wan team's clip models did go open, and the authors have said they hope to do the same here, but a research cadence this fast with no release is worth reading as a research programme rather than a product path. Treat it as the clearest picture available of where real-time AI companions are heading, and not as something you can build on this quarter.

The other direction Wan moved in July 2026 was down. Qualcomm AI Research published MobileWan, which runs the 5B-parameter Wan 2.2 architecture entirely on commercial Snapdragon mobile hardware. It generates 81 frames, five seconds at 16fps, at 480x832, in roughly 20 seconds end to end on the device itself. The whole download is about 9.9 GB. Note the attribution, because plenty of coverage got it wrong: this is Qualcomm's engineering built on Alibaba's open weights, not an Alibaba release.

The compression work is what makes it possible. MobileWan generates in chunks rather than processing every frame together, prunes the attention heads that contribute least, uses a memory-efficient decoder, and cuts diffusion from the usual 20 to 30 steps down to three. That last number is the one doing the heavy lifting on latency.

Quality is not competitive with server-scale Wan, and it is not trying to be. What it changes is the floor. Generation on your own phone means no account, no upload, no queue, and no hosted content filter between you and the output, which is the combination self-hosters have wanted for years and previously needed a desktop GPU to get. It also breaks the cost model our video cost guide is built on: when generation happens on hardware you already own, the marginal cost of a clip is battery rather than API credits.