A filmmaker in Hangzhou used to wait three weeks and a five-figure budget for a single 10-second visual-effects shot. Last month she generated twelve variants of the same shot before lunch, picked one, and shipped it. The tool that closed that gap was not a Hollywood suite but Kling AI (可灵AI), the video-generation model spun out of Kuaishou (快手), the short-video giant with 741 million monthly active users.
Kling is no longer a lab demo. In Kuaishou's 2025 full-year results, the model posted commercial numbers that few generative-video products anywhere can match.
Key takeaways
- Kling AI's December 2025 monthly revenue exceeded US$20M, lifting its annualized run-rate (ARR) to US$240M (≈ HK$1.86B).
- Fourth-quarter 2025 revenue reached RMB 340M (≈ US$48M / HK$370M).
- By the end of 2025 the product had over 60 million global users, more than 600 million videos generated, and over 30,000 enterprise and developer API customers.
- Three model generations shipped in roughly a year: Kling O1 (unified multimodal), Kling 2.6 (audio+video in one pass), and Kling 3.0 (Feb 2026, All-in-One full-modal).
From silent clips to a working studio
The story of Kling's 2025 is the story of collapsing the production pipeline. Early text-to-video tools produced a few seconds of wordless footage; everything afterward — voice, sound effects, character consistency — was still hand-assembled. Kuaishou attacked that seam directly.
Kling O1, launched in the fourth quarter of 2025, was positioned as the first unified multimodal video model: it fuses text, video, image and subject references into a single engine so a creator never has to jump between generation and editing tools. Kling 2.6 added "音画同出" (synchronized audio-visual output) — the model emits a finished clip carrying natural dubbing, action sound effects and ambient tones in one pass, instead of generating silent footage and dubbing it later. It also shipped motion control: upload a reference video or pick an action from a library, add a subject image, and the model reproduces the pose and expression on that character.
Kling 3.0, released in February 2026, takes the All-in-One (All-in-One) philosophy further. It supports full-modal input and output across text, image, audio and video, folding understanding, generation and editing into one workflow. Kuaishou says it can follow complex narrative logic, auto-storyboard, and hold precise camera control with high prompt adherence. Some Chinese coverage bills the "图生视频 + 主体参考 (image-to-video plus subject reference)" capability as a global first; that is the company's own framing and should be read as a claim, not an independently verified benchmark.
Why the revenue matters more than the demos
Generative video has a long history of impressive clips and weak balance sheets. Kling's numbers invert that. A US$240M run-rate inside a consumer-internet company is notable because it proves a willingness to pay, not just a willingness to watch.
Two forces explain it. First, Kuaishou embedded Kling directly into commercial workflows — marketing, e-commerce, film, microdrama (短剧), animation and gaming — rather than leaving it as a playground. Second, the same infrastructure that serves 741 million monthly users gives Kling a distribution and inference-cost advantage most standalone startups lack. On the company's earnings call, CEO Cheng Yixiao (程一笑) said Kling's ARR had passed US$300M by January 2026 and expressed confidence in more than doubling revenue in 2026. That forward comment is a management target, not a booked result. Crucially, the revenue is recurring: API access, enterprise seats and creator subscriptions, not one-off hype. That is the difference between a model that trends for a week and one that funds its own compute.
Where the paying users actually are
Kling's run-rate is not spread evenly across China. Company statements and Chinese media coverage indicate that more than 70% of its revenue comes from outside the mainland — North America, Europe, Japan, South Korea and Southeast Asia — where creators and small studios pay for credits rather than treat the tool as a free toy. That overseas tilt is a deliberate hedge: it keeps Kling clear of the crowded domestic short-video-generator market while riding a window when Western rivals like Sora were slow to open commercial tiers.
A concrete slice of that demand is microdrama (短剧). Kling's reference-and-motion workflow lets a small studio generate dozens of consistent character variants for a 100-episode vertical drama without a film crew, and the company has pointed to adoption by film and television pipelines as a core revenue vertical alongside e-commerce and gaming. By mid-2026 Kuaishou said Kling's global users had passed 100 million and its enterprise and developer customers neared 50,000 — a sharp step up from the 60 million users and 30,000 enterprises it reported at the end of 2025.
What readers can do now
- If you produce short-form video, test Kling's reference-and-motion features before budgeting a reshoot. The motion-control and subject-reference workflow is where the cost gap versus traditional visual effects is widest.
- Enterprise teams should price the API, not the demo. With 30,000+ developer and enterprise customers already on board, the integration path is mature enough to model into a real pipeline.
- Watch the ARR trajectory, not the clip quality. Kling's competitive moat is commercial distribution; track whether the US$300M-by-January claim holds in the next earnings cycle.
Honest limitations
- All revenue and user figures come from Kuaishou's official 2025 results and earnings-call commentary, reported by multiple financial outlets; they are company-disclosed, not independently audited by a third party.
- The "global first" and capability-supremacy claims around Kling 3.0 originate from Kuaishou and Chinese media; I did not find a head-to-head independent benchmark confirming them at publication time.
- The US$300M ARR-by-January-2026 statement is a management aspiration stated on an earnings call, not a finalized reported number.
- This article covers Kling AI only; it is not an endorsement and does not rank every competing model on identical tasks.