Writing a sentence rarely produces a film. For years, text-to-video tools could fake a few pretty seconds but fell apart on anything with a plot, a camera move, or a face that had to stay the same. Kuaishou, the short-video giant behind Kling AI, just pushed its base model into a new generation — and the interesting move is less the prettier frames than how it lets creators talk to the model.
Kling AI 2.0 is the latest iteration of a large model (大模型) for video and image generation that has quietly become one of the most used creative AI tools in the world.
From vision to screen
Kling AI 2.0 was unveiled at a Beijing launch event themed "From Vision to Screen." Two models moved up together: the Kling AI 2.0 Video Generation Model and the Kolors 2.0 (可图) Image Generation Model.
Kuaishou positions Kling as the world's first user-accessible DiT (Diffusion Transformer) video model. Since its original launch in mid-2024, the system has gone through more than 20 iterations focused on image clarity, motion quality and new creative controls.
The scale of usage is the part that surprises outside observers:
- 22 million+ global users since the mid-2024 launch.
- 15,000+ developers and enterprise customers wiring Kling's API into their own products.
- 168 million videos and 344 million images generated on the platform to date.
- A 25× growth in monthly active users across the first ten months.
Those are Kuaishou-reported figures from the launch, but they point to a model that crossed from novelty to daily tool faster than most lab releases.
The benchmark that matters
Independent validation is rarer and more useful. On March 27, the AI benchmarking group Artificial Analysis published its global video-generation ranking, where Kling 1.6 Pro (high-quality mode) took first place in the Image-to-Video category with an Arena ELO score of 1000, ahead of Google's Veo 2 and Pika.
On its own internal GSB (Good-Same-Bad) tests, Kuaishou claims Kling 2.0 beats Google Veo 2 and Runway Gen-4 on semantic responsiveness, visual quality and motion quality — reported win-loss ratios of roughly 182% against Veo 2 and 178% against Runway Gen-4. Treat those internal numbers as vendor claims, but the independent Arena ELO result is a genuinely external data point.
The real innovation: talking to video with MVL
The headline feature of 2.0 is not just better output — it is a new way to direct the model. Kuaishou introduced Multi-modal Visual Language (MVL), an interaction concept built from two pieces:
- TXT — pure text as the semantic skeleton of a shot.
- MMW (Multi-modal-document as a Word) — image references, video clips, and eventually voice or motion trajectories, used as direct instructions.
The idea is that words alone are a poor way to describe a shot. A creator can now hand the model a reference image for identity, a clip for style, and notes for camera movement, and Kling assembles them. The "Master Edition" adds multimodal editing so users revise existing footage by showing, not just describing.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the shift from text-only prompts to multimodal direction is where Chinese video models are genuinely pulling ahead — the bottleneck was never the pixels, it was the conversation.
Why this is a foundation-model story
Kling is easy to file under "consumer app," but it is fundamentally a foundation-model release: a large generative model, iterated rapidly, exposed through an API, and benchmarked against the best closed systems. Its trajectory — open benchmark leadership plus mass consumer adoption — is the template other Chinese model labs are racing to copy.
These models are part of a wider Chinese push that also spans humanoid (人形机器人) robotics and brain-computer interface (脑机接口) research, but Kling's contribution is in making generative video both good enough and controllable enough for real production.
Honest limitations
- Vendor-heavy metrics. The 22M users, 15K developers and GSB win-rates are Kuaishou's own; only the Artificial Analysis Arena ELO is independent.
- Still a creative tool, not a director. Long, coherent narrative video with consistent characters across minutes remains hard; MVL helps control, it does not guarantee story.
- Access and cost vary by region. API pricing, quotas and content policies differ across markets, so "studio-grade" depends on which tier you can use.
- Benchmark leadership moves fast. Video-generation rankings shift quarterly; a #1 today is not a #1 next spring.
What readers can do now
- Test the Arena ELO claim yourself. Run the same image-to-video prompt through Kling and a competitor and judge consistency of face, motion and prompt adherence.
- Try MVL, not just text. Upload a reference image as an MMW and compare the control you get versus a pure-text prompt — the gap is the real story.
- Watch the API for production use. If you build creative tooling, benchmark Kling's API on your own assets before committing, since public demos rarely reflect enterprise workloads.