A filmmaker feeds the model a single product photo and a line of text. A few seconds later, a short, coherent clip exists — the object turning, the light shifting, the background holding steady the whole time.
No render farm. No crew. Just weights anyone can download and run.
That is the bet behind SkyReels, a family of video models from Kunlun Tech, one of the Chinese companies betting that the future of video generation belongs in open repositories, not behind paywalls.
What SkyReels actually does
In late January 2026, Kunlun's Skywork AI team released SkyReels-V3 as open-source code. The model is not a single trick; it ships three capabilities that each solve a different creator headache.
- Reference image-to-video: feed it one to four reference images plus a text prompt, and it generates a time-consistent clip that keeps the subject's identity, the spatial layout, and the story intact.
- Video extension: hand it an existing short clip and it continues the scene — not just by stretching time, but by predicting the next sensible shot, including real editing transitions like cuts and reverse shots.
- Audio-driven talking avatar: from one portrait and an audio track, it produces a lip-synced, expressive speaking character (the company frames this as one module among several, not the headline feature).
The first two capabilities are the ones most relevant to everyday creators, because they turn a still or a fragment into usable footage without a studio.
Why giving it away matters
Open-sourcing a video model is a strategic choice, not a charity. Kunlun has released a string of SkyReels versions — V1, V2, the A1, A2, and A3 spin-offs, then V3 — and the company says the family has pulled close to 300,000 cumulative downloads on Hugging Face with more than 10,000 GitHub stars. Those are vendor-reported adoption numbers, so read them as evidence of community reach rather than audited usage.
The point of open release is the ecosystem it seeds. Developers fork the code, tune it for a niche, file issues, and ship derivatives. That feedback loop is hard for a closed product to match, and it plants the model inside other people's pipelines — exactly where long-term dependence forms.
Kunlun Tech is listed in Shenzhen as 300418.SZ, and the SkyReels work comes out of its Skywork AI research group.
The benchmark that got attention
A month after V3 opened up, Kunlun released SkyReels-V4 in February 2026. In its annual report filing, the company states that V4 ranked first on Artificial Analysis's text-to-video (with audio) and image-to-video (with audio) leaderboards at launch, ahead of models from Kling, Google's Veo, and OpenAI's Sora line.
That claim comes from Kunlun's own regulatory disclosure, not from an independent study we verified separately, so it should be treated as the company's stated result rather than settled fact. Leaderboards also move constantly; a top spot in February is a snapshot, not a permanent crown. What is verifiable is that a Chinese open model reached the front of a recognized external benchmark at all — that alone is a marker of how far the open-video field has come.
Where the large model fits
SkyReels is built on a diffusion-transformer large model (大模型) architecture, the same broad family that powers much of today's generative AI (生成式人工智能) video work. Its differentiator, per Kunlun, is holding identity and motion consistent across longer, multi-shot sequences — the part that usually breaks down when a clip has to tell a story instead of showing a single effect.
The company has also opened an API for V4, so teams that do not want to host the weights can call the model directly. That dual path — open weights for tinkerers, hosted API for businesses — is becoming the standard playbook for model labs that want both community and revenue.
What creators can build with it
The practical uses are narrower and more useful than "make a movie."
- An e-commerce seller drops a product shot into the reference mode and gets a rotating, styled demo clip without a photo studio.
- A short-video creator extends a good take into a second scene instead of reshooting.
- A small studio fine-tunes the open weights for a consistent character across a series, then deploys that character through the API.
None of this replaces a production pipeline for a theatrical film. But for the enormous volume of commercial and social content made on thin budgets, an open, downloadable engine changes who gets to produce at quality.
Honest limitations
The download and GitHub-star counts are Kunlun's own reported figures from its releases and filings, not independently audited usage data, so they indicate reach rather than proof. The Artificial Analysis ranking claim appears in Kunlun's annual report and has not been re-verified by us against the live leaderboard, and video leaderboards shift week to week. SkyReels-V3's talking-avatar module is a real capability but was not the focus of this piece, and running these models locally still demands meaningful GPU resources despite the open release. This article is current as of 22 October 2026 and is not investment advice.
What readers can do now
- Pull SkyReels-V3 from Hugging Face and run the reference-image-to-video mode on one of your own assets to see how well it keeps identity across a short clip.
- If you ship video products, prototype the open-weights-plus-API split: build on the free weights, then call the hosted API in production.
- When you see a "number one" video claim, check the leaderboard date and whether audio was included before repeating it.