Alibaba

Wan 2.7

the open workhorse

3 min readVideo Generation

Key facts

Apache 2.0permissive
Licence
1.3BT2V params
Small variant
Localconsumer GPU
Runs on
Wanseveral sizes
Family
2.7Alibaba
Version

The open workhorse. Apache 2.0, part of Alibaba's open Wan family, with small variants including a T2V-1.3B that runs on consumer hardware for concepting loops.

What it is

Wan 2.7 is Alibaba’s open video model, and it has become the workhorse of the open-source end of this section. Released under the permissive Apache 2.0 licence and part of Alibaba’s wider Wan family, it comes in several sizes, including a compact T2V-1.3B variant that is 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 has made Wan 2.7 the default choice for people who want to generate video without paying for an API.

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.