Open-Source Video Models Mid-2026: What Actually Shipped, and What Did Not

This is the fourth entry in our open-source AI video series. In the April comparison we tested Wan 2.1 and 2.2, HunyuanVideo and Mochi on real hardware and named a winner per use case. Three months on, the obvious follow-up was "here are the new models." We started writing that piece and the facts did not cooperate.

The headline change since April is not a new open model. It is that the web filled up with confident, well-ranked claims about a model that was never opened at all.

TL;DR

• Wan 2.7 is not open source. Official Wan open weights still stop at 2.2. There is no Wan2.7 repository and no weights on Hugging Face, despite a cluster of sites saying otherwise.
• LTX-2.3 is the real story. 22B parameters, synchronized audio and video from one model, and over 2.1 million downloads. It is the most-downloaded open video model we can find by a wide margin.
• None of the leaders is Apache 2.0 at the frontier. The "they all went permissive" claim is wrong in three different ways, laid out below.

The Wan 2.7 problem

Search for open-source video models today and you will be told, repeatedly and confidently, that Wan 2.7 shipped under Apache 2.0 in April 2026 with first/last-frame control and 9-grid image input. Several of those pages give install instructions and download links.

We went looking for the weights. They are not there:

Wan 2.7 does exist. It is a real model with real capabilities, available through Alibaba's cloud and API. What it is not is downloadable. Alibaba's pattern with Wan has been cloud first, open weights later, and for 2.7 the second half has not happened.

The sites saying otherwise share a tell: many are single-purpose domains built around the model name, publishing "open source guide" and "how to download" pages for something with nothing behind them. They rank well because nobody else was competing for those terms.

The practical consequence for readers of our April piece is reassuring. We tested Wan 2.1 and 2.2 because that was the open Wan, and 2.2 is still the open Wan. That article did not age out. The newer content did.

LTX-2.3 is what actually shipped

The model that genuinely changed the picture came from Lightricks. LTX-2 was open-sourced in January 2026; the current release is LTX-2.3, a 22-billion-parameter diffusion transformer that generates video and synchronized audio together from a single model rather than bolting a soundtrack on afterwards. It handles text-to-video, image-to-video, and video-to-video.

The adoption numbers are what stand out. Set against the other open contenders:

Downloads measure interest rather than quality, and the Wan totals are split across many repository variants, so do not read this as a capability ranking. But a 500x gap over HunyuanVideo-1.5 is not noise, and it lines up with the other signal: an active adapter ecosystem. Lightricks has shipped LoRAs through June and July for video-to-audio foley, cinemagraphs, in/outpainting, clean-plate removal, and spatial upscaling. That is what a model people actually build on looks like.

Requirements

LTX-2.3 wants Python 3.12 or newer, CUDA above 12.7, and PyTorch 2.7 or later. Resolution has to be divisible by 32, and frame count divisible by 8 plus 1, which will bite you the first time you request 100 frames and get an error.

The licensing claim, corrected

The tidiest version of the mid-2026 story is that the open video leaders converged on Apache 2.0. We were ready to write that. It is wrong three times over:

The LTX threshold has a detail worth reading before you assume you are under it: subsidiaries, affiliates, and companies under common control are counted collectively. A small studio owned by a large group may be over the line. Lightricks states there are no retroactive charges, so crossing the threshold is a conversation rather than a penalty.

For most readers of this site, "free under 10 million ARR" is functionally free. It is still not Apache 2.0, and if you are choosing a model on license terms the difference matters.

Check any model claim yourself in about a minute

The useful takeaway is not the specific facts above, which will move again. It is that both Hugging Face and GitHub expose public APIs that settle "is this actually released" in one command, with no key and no account.

Does the model exist on Hugging Face?

curl -s "https://huggingface.co/api/models?search=LTX-2.3&limit=5" \
  | python3 -c "import sys,json; [print(m['modelId'], m.get('downloads')) for m in json.load(sys.stdin)]"

An empty result is your answer. That single check is what showed us Wan 2.7 had no weights.

What has an organization actually published, most recent first?

curl -s "https://huggingface.co/api/models?author=Wan-AI&sort=lastModified&direction=-1&limit=10" \
  | python3 -c "import sys,json; [print(m['modelId'], str(m['lastModified'])[:10]) for m in json.load(sys.stdin)]"

And the repository's real license, rather than what a blog says it is:

curl -s https://api.github.com/repos/Wan-Video/Wan2.2 \
  | python3 -c "import sys,json; j=json.load(sys.stdin); print(j['license']['spdx_id'], j['pushed_at'])"

A NOASSERTION here means GitHub could not match the LICENSE file to a standard license, which is a strong hint you are looking at a custom community license rather than Apache or MIT. That is exactly what HunyuanVideo-1.5 returns.

What this changes about the April comparison

Less than you would expect, which is the point.

Our tested numbers for Wan 2.2 and HunyuanVideo still describe the current open field, because the open field did not move as much as the coverage suggests. The genuine gap is LTX-2.3, which we have not benchmarked and which did not appear in that article at all. If you want the hardware side, the self-hosted stack guide still applies; a 22B model sits above the Wan 2.2 5B variant and below the heavier A14B configurations in memory terms.

What we would test next

Three things for the next entry. LTX-2.3 against Wan 2.2 on the same prompts and hardware as the April run, so the comparison is like for like. Whether the single-model synchronized audio actually beats generating video and audio separately, which is LTX's core architectural bet. And whether the foley and upscaler LoRAs hold up outside the demos, since an adapter ecosystem is only worth something if the adapters work.

If Alibaba does open Wan 2.7, that becomes the story and we will cover it. It has not happened yet.

Sources and further reading

• Part one: our tested open-source video model comparison
• The self-hosted video stack hardware guide
• Lightricks/LTX-2.3 on Hugging Face
• The LTX-2 Community License Agreement
• Wan-Video/Wan2.2 on GitHub (Apache-2.0)
• Tencent HunyuanVideo-1.5 on GitHub

Tested on: not independently tested. We did not run LTX-2.3, which needs an NVIDIA GPU our bench does not have. Every release claim above was verified directly against the Hugging Face and GitHub public APIs on the date below, using the exact commands published in this article: model existence and download counts, organization publishing history, repository licenses, and last-push dates. Capability descriptions are the projects' own. Where our findings contradict widely-published claims, we have shown the check so you can repeat it.
Date checked: 2026-07-25