Wan 2.7

What it is

Wan 2.7 is the generation of Alibaba’s Wan video family covered here, the workhorse line of the open-source end of this section. A precision that most write-ups skip, and the state of play recorded in early August 2026: Wan 2.7 itself was available through hosted APIs, while the weights openly published on Alibaba’s own repositories stopped at Wan 2.2, released under the permissive Apache 2.0 licence in several sizes, including a compact variant small enough to run on consumer hardware. That combination of a free licence and a model that fits on an ordinary graphics card is what made the Wan family the default choice for people who want to generate video without paying for an API, and the thing to check before self-hosting is which generation the downloadable weights have reached.

The model family

The family structure is the first thing to understand. Rather than shipping a single monolithic model, Alibaba releases Wan 2.7 as a range of sizes, from larger variants aimed at quality to the 1.3-billion-parameter text-to-video model built for reach. The small model is the interesting one. At 1.3 billion parameters it will run on the kind of graphics card a hobbyist or a small studio already owns, which means a creator can generate short clips locally, in a tight loop, without sending anything to a server or paying per generation. Alibaba positions that variant for concepting loops: quick, cheap, repeatable passes where the point is to try an idea rather than to produce a finished frame.

Why workhorse fits

That local, low-cost workflow is why the word workhorse fits. A model that runs on your own hardware can be used all day without a bill mounting up, and it can be woven into a pipeline in ways a metered endpoint cannot. For iteration, where a creator might generate dozens of variations before settling on one, the economics of a free, locally hosted model of this kind are hard to argue with. The larger variants in the family cover the cases where output quality outranks cost, so the range as a whole spans from fast concepting to more polished delivery.

The strategic picture

The strategic picture is the part worth stepping back for. The open Chinese video stack now mirrors the open Chinese approach to large language models: permissive licences, competitive quality, and a clear intent to make paid Western tools look expensive. Wan 2.7 is a central piece of that. By putting a capable video model under Apache 2.0 and making the smaller versions genuinely runnable on consumer hardware, Alibaba applies the same pressure to the video market that its open language-model releases applied to the language-model market, driving the price of a baseline capability towards zero and forcing the paid tools to justify their premium on quality and integration alone.

That pressure changes the calculation for everyone else in the field. A Western startup weighing whether to build on a closed API now has a credible free alternative it can host itself, and the closed vendors have to price and position against a model that costs nothing to license. The model is not the flashiest in this section, and it does not need to be; its role is to be good enough, free, and everywhere, which is a more disruptive position than being briefly best. Because the weights are open, a community of tool-builders can wrap, fine-tune and optimise the model without asking permission, and that unpaid ecosystem work tends to compound over time in a way a single vendor cannot match on its own.

Where it sits

For anyone mapping the open end of the AI video field, Wan 2.7 sits alongside a small group of genuinely open models as the workhorse option, the one to reach for when you want capability without a contract. The parallel with the open language-model story is the thing to watch: if video follows the same path, the base level of quality will keep rising while the price of access keeps falling, and Alibaba’s Wan family will have been one of the reasons why. For the wider context, see the AI hub.