Kling — The Commercial Video Model That Made Physics Simulation a Product Feature

When Runway and Pika were defining what “AI video” meant to Western audiences in early 2024, a different kind of video generation system was being built inside Kuaishou Technology in Beijing. It was not a research project. It was not a demo. It was a commercial product being engineered by one of China’s two largest short-video platforms — a company with 24,000 employees, a Hong Kong-listed stock, and the operational experience of encoding and serving billions of videos per day to hundreds of millions of users.

Kling launched on June 10, 2024, via Kuaishou’s investor relations channel. It reached USD 100 million in annualized recurring revenue in its tenth month. By December 2025 — nineteen months after launch — it had crossed USD 240 million ARR, 60 million registered creators, and 600 million videos generated. Kling 3.0, released February 2026, generates native 4K video at up to 60 fps with synchronized audio and multi-shot storyboarding in a single pass.

This is not a research artifact or a community curiosity. Kling is a full-scale commercial AI video platform, and by any measure it is the most commercially successful open-access video generation system ever built by a Chinese company. Understanding what it does well, what it does not, and what it means for the landscape of AI video requires looking at both the technical choices and the business context that shaped them.

We write from public sources — Kuaishou investor relations, arXiv technical reports, independent press coverage, community evaluations, and pricing documentation. We do not test AI video models hands-on.

Background: Kuaishou Technology and the Platform That Built Kling

Kuaishou was founded in 2011 by Hua Su and Cheng Yixiao in Beijing, originally as a tool for sharing animated GIFs. By 2013 it had pivoted to short-form video. By 2016 it had added livestreaming and was growing rapidly as China’s first large-scale short-video platform — several years before TikTok’s international expansion would introduce that format to Western markets. Kuaishou IPO’d on the Hong Kong Stock Exchange in February 2021 (ticker: 1024.HK), briefly hitting a market capitalization of over USD 159 billion on its debut day.

Today Kuaishou is China’s second-largest short-video platform, behind ByteDance’s Douyin/TikTok. Revenue runs approximately 124 billion CNY (~USD 17 billion) annually, split across advertising, livestreaming gifts, and e-commerce commissions. The international app is branded Kwai (briefly called Snack Video in South and Southeast Asian markets before India banned it in 2020 alongside a broad set of Chinese apps).

Kling is Kuaishou’s entry into the generative AI layer of that business. The logic is straightforward: a platform that processes, stores, and serves video at planetary scale is well-positioned to build tools that generate video. The infrastructure investment, the model training compute, the data flywheel from billions of user-created clips — all of it compounds.

The announcement of Kling came with a specific framing. Kuaishou did not position it as a research contribution or an open-weight release. It was a commercial product being offered to creators who use Kuaishou’s tools, and then to the global market via klingai.com and the international kling.ai interface.

Architecture: Diffusion Transformer Plus a 3D VAE

Kuaishou has been more transparent about Kling’s architecture than most commercial video AI labs. The publicly documented technical picture:

Backbone: Diffusion Transformer (DiT)

Kling uses a Diffusion Transformer architecture — the same family as Sora (reported), CogVideoX, and most recent high-quality video models. This is a meaningful choice: DiT architectures scale well with compute and have demonstrated qualitatively different motion quality at large scale compared to earlier U-Net-based approaches like AnimateDiff or Stable Video Diffusion.

3D VAE

The core technical differentiator Kuaishou highlighted at launch was their self-developed 3D Variational Autoencoder, which performs synchronous spatiotemporal compression of video frames. Standard video VAEs (inherited from image diffusion systems) compress each frame independently, then process the sequence in a temporal attention layer. Kuaishou’s 3D VAE treats spatial and temporal dimensions jointly during compression, enabling higher reconstruction fidelity at a given token count and supporting the kind of temporally coherent latent representation that makes fluid and cloth motion tractable.

Full Spatiotemporal Attention

Rather than the factorized attention approach (separate spatial and temporal attention passes) used in earlier architectures for computational tractability, Kling implements full joint spatiotemporal attention — spatial and temporal tokens attending to each other simultaneously within each transformer layer. Kuaishou describes this as enabling the model to capture both local per-frame features and cross-frame motion dynamics in a unified representation, rather than stitching together the two in post-processing.

No formal paper for Kling 1.x

The original launch was documented via investor relations press releases and a product blog, not an arXiv preprint. The most technically detailed public disclosure came later.

Kling-Omni Technical Report (arXiv: 2512.16776, December 2025)

With the O1 launch, Kuaishou published a proper technical report. The Omni architecture is described as a unified diffusion transformer processing visual and textual tokens in a shared embedding space — similar in spirit to the multimodal joint attention approaches in other recent large models. The training pipeline involves four stages: large-scale pretraining on text-video pairs; supervised fine-tuning on multimodal curriculum tasks (reference-to-video, editing); Direct Preference Optimization (DPO) for motion dynamics and visual integrity; and model distillation that reduces inference from 150 to 10 neural function evaluations. Infrastructure includes 3D parallelism, multimodal FlashAttention, FP8 quantization, and reference-aware KV caching that provides approximately 2× inference speedup.

Version History: Eighteen Months of Rapid Iteration

No commercial AI video model has iterated as publicly and consistently as Kling over its first nineteen months. Understanding the version trajectory matters because the model that exists now is substantially different from what launched.

Kling 1.0 — June 2024

The initial version offered text-to-video and image-to-video generation up to 1080p at 30 fps, with video extension up to 2 minutes. This was the announcement — technically impressive for a first commercial release, but without the specific control features that would define the product.

Kling 1.5 — November 2024

The update that made Kling a serious contender for professional workflows. 1.5 introduced:

  • Motion Brush: Paint directional motion trajectories onto up to six independent regions of an image; designate other regions as static. This level of per-element motion control was not available in other commercial systems at the time.
  • Camera movement controls: Six named presets — horizontal pan, vertical pan, zoom, pan, tilt, roll — providing editorial-grade control over camera behavior without manual specification.
  • Standard and Pro quality modes: 720p Standard and 1080p Pro.
  • Video Extension to 3 minutes total: Extend generated clips by 4–5 seconds incrementally.

Motion Brush in particular drew attention in the production community. Being able to specify “the water moves left, the trees move slightly, the background is static” without re-prompting or inpainting is a qualitatively different kind of control than anything available in the then-current Runway or Pika interfaces.

Kling 1.6 — December 2024

Iterative quality improvements. 1.6 Standard (720p, 5 seconds, 24 fps T2V) and 1.6 Pro (1080p, up to 10 seconds, with first-frame and last-frame conditioning). The last-frame conditioning — generating a clip that transitions from a specific start image to a specific end image — was a genuinely novel capability that competitors lacked at the time.

Kling 2.0/2.1/2.5 — April–Mid 2025

The 2.x series focused on quality consolidation: improved physics simulation, better lighting, stronger prompt adherence, start-and-end keyframe control (2.1), and a Turbo speed variant (2.5) for faster iteration workflows. By mid-2025 Kling 2.5 was widely regarded as peer with Runway Gen-3 Alpha on photorealism, while maintaining its advantage in motion physics.

Kling 2.6 — December 2025

A significant technical milestone: synchronized audio-visual generation in a single diffusion pass. Not post-processing dubbing applied on top of a generated video — audio and video generated together from a shared latent representation, resulting in lip-synced dialogue, synchronized sound effects, and contextually appropriate ambient audio. Resolution improved to 1080p at 48 fps. This capability arrived before Runway or Pika offered native audio co-generation.

Kling O1 / Omni — December 2025

The architectural leap. Kling O1 introduced the Unified Multimodal Video Model (documented in arXiv:2512.16776): a single model engine handling text-to-video, image-to-video, video editing (in-painting, content insertion/removal), style re-rendering, attribute manipulation, and reference-based generation with multi-image identity libraries. Rather than separate fine-tuned models for each task, O1 routes all generation tasks through a shared transformer with task-conditional routing. Digital Human 2.0 — for long-duration, identity-consistent talking-head avatar generation — launched the same week.

Kling 3.0 — February 4, 2026

The current version at the time of writing. Key capabilities:

  • Native 4K (3840×2160) at up to 60 fps — the first major commercial T2V model to offer 4K in a production API
  • Up to 15 seconds per generation in a single pass (up from 10 seconds)
  • Multi-shot storyboarding: Up to 6 distinct camera cuts per generation, with character and style consistency maintained across all cuts. This is a departure from the single-shot paradigm — rather than generating a clip, you can describe a scene that includes multiple camera angles and transitions.
  • Native audio in 5 languages: English, Chinese, Japanese, Korean, Spanish; dialogue generation with accent and dialect support, integrated music and SFX
  • Character cloning: Reference-based generation with consistent character identity across shots

What Kling Is Known For

Physics Simulation

The claim that appears most consistently across independent comparisons of commercial video models is that Kling leads on physics. Fluid dynamics — water, liquid, fire, smoke — consistently receive praise for temporal coherence that other models fail to match. Fabric and cloth simulation, particularly billowing or draping material, is similarly strong. Human body movement produces natural articulation rather than the sliding-through-space artifacts common in architectures that lack explicit temporal modeling.

This is not a subjective impression. The community evaluations that have tracked Kling since 1.5 consistently place it at or near the top for physical plausibility. Kuaishou’s internal benchmark (OmniVideo-1.0, 500+ evaluation cases) claims superiority over Google’s Veo 3.1 and Runway Aleph on motion dynamics — though these are self-reported figures, not third-party evaluations.

Motion Brush and Camera Controls

Motion Brush (1.5+) remains one of the more distinctive control features available in any commercial video system. The ability to specify independent motion trajectories for different image regions — this element moves up-left, that element moves slowly right, the background stays still — while preserving visual coherence across the frame, enables a level of creative direction that purely prompt-based systems cannot match.

The camera movement presets are table stakes by 2026, but Kling’s implementation has been consistent and reliable since 1.5. For directors who think in terms of named camera moves, having dedicated controls rather than trying to describe camera motion through text is a meaningful workflow improvement.

Start/End Frame Keyframing

Since Kling 1.6 Pro, the ability to specify both start and end frames gives filmmakers control over arc rather than just origin. Generate a character standing at the start of a clip who is seated at the end. Generate a landscape at dawn that reaches golden hour. This kind of temporal control over a clip’s narrative arc is harder to achieve through prompting alone.

Commercial Scale and Reliability

By February 2026, Kling has served over 600 million generated videos. The infrastructure at that scale is materially different from a research demo or a small-scale commercial API. Generation times are consistent, the system handles production loads, and the version cadence suggests an engineering organization that is resourced to maintain and extend the product.

Access, Pricing, and Distribution

Kling is closed-source and commercial. There is no open-weight release and none has been announced.

Consumer web interface: kling.ai (international), klingai.com (China and developer API)

Free tier: 66 credits per day with rollover. Free-tier generations are lower resolution (360p–540p), watermarked, and generated in a public queue — prompts and reference images may be visible in the community gallery. This is a meaningful privacy consideration for professional work.

Subscription plans (approximate mid-2025 structure; pricing subject to change):

Tier Price/month Credits/month Notes
Standard ~$10 660 Up to 720p; ~33 standard-length videos
Pro ~$37 3,000 720p–1080p; Private Mode enabled
Premier ~$92 8,000 1080p; ~400 standard videos
Ultra ~$180 26,000 1080p+; high-volume production

Annual billing reduces cost approximately 34%. Private Mode — preventing generations from appearing in public galleries — requires Pro tier or above.

Enterprise API: The official Kling developer API (klingai.com) is structured around substantial upfront commitment — roughly USD 4,200 for a 3-month package providing approximately 10,000 credits per month. At that rate, a single 10-second Pro-mode video costs approximately $1 via the official API. Third-party API resellers (PiAPI and others) offer flexible, lower-commitment access at comparable or slightly higher per-generation pricing.

Third-party platform availability: Kling is available embedded in Pollo AI, Scenario, Atlabs, fal.ai, Higgsfield, and other multi-model platforms. For users who already use these platforms, Kling access may require no separate subscription.

How Kling Compares

By early 2026, the commercial T2V landscape has consolidated around a small group of systems capable of native 4K with synchronized audio: Sora 2, Veo 3.1, Kling 3.0, Seedance 2.0, and Runway Gen-4.5. Kling sits clearly in that tier. Within it:

vs Runway Gen-3/Gen-4: Runway leads on precision camera controls and professional editorial output quality, and is the dominant choice for commercial film and advertising production. Kling leads on physics simulation and value for volume creation. Runway Gen-4.5 is generally rated the top closed-source professional tool for visual fidelity per independent leaderboard data. At comparable output quality, Kling’s per-second cost is substantially lower than Runway.

vs Sora: Sora delivers smoother motion with stronger narrative coherence across longer clips. Kling’s physics simulation is widely considered comparable or superior; Kling generates faster and is more accessible internationally (Sora requires ChatGPT Plus). The content moderation constraints on Kling (see below) are a meaningful practical difference for some use cases.

vs Luma Dream Machine: Luma excels at cinematic camera motion — smooth tracking shots and dolly moves that feel cinematographically intentional. Kling excels on subject/character consistency and motion control precision. They are roughly competitive on raw visual quality; the choice often comes down to whether you need camera smoothness or physics fidelity.

vs Pika: Pika 2.5 trails Kling on raw realism and physics. Pika’s Pikaframes (start/end frame control, introduced before Kling’s equivalent) remains a standout feature, and the pricing is generally lower. For high-fidelity physical simulation, Kling is the better tool.

Kling is a product of a Chinese company subject to Chinese law, and that has practical consequences.

Content filtering: Kuaishou’s models are required by the Cyberspace Administration of China (CAC) to “embody core socialist values.” In practice, this means prompts referencing the Tiananmen Square protests, criticism of Xi Jinping, calls for Taiwan independence, and other politically sensitive topics will be rejected with error messages. This is not incidental to the product — it is a regulatory requirement for Chinese AI models. TechCrunch documented specific instances of politically sensitive prompts being silently filtered as early as July 2024. For most commercial creative applications this constraint is irrelevant; for journalistic, documentary, or activist use cases it is a material limitation.

Data handling: Kuaishou’s privacy policy grants a worldwide, non-exclusive, royalty-free, sublicensable license to use, store, reproduce, and modify user content. Processing may occur on servers in China. The free tier’s public queue means free-tier prompts and reference images may be visible to other users. Privacy-sensitive work — featuring real people, confidential products, unreleased creative projects — should use at minimum the Pro tier with Private Mode enabled.

No US sanctions or export controls: Unlike Nvidia GPU hardware exports, generative AI SaaS services from Chinese companies are not currently subject to formal US export controls. Some enterprise security policies or government contractor contexts may have internal restrictions on using Chinese AI services, but there is no broad legal prohibition on US users accessing Kling.

Phishing risk: Check Point Research documented a malware campaign in early 2025 using fake Kling AI websites to distribute infostealer malware. This is not a flaw in Kling itself, but it is a sign of the platform’s brand recognition being large enough to be impersonated. Access Kling only through kling.ai and klingai.com.

Commercial Trajectory

The business numbers are unusual for an AI product in this category.

Kling reached USD 100M ARR in March 2025 — ten months after launch. It crossed USD 240M ARR in December 2025, with monthly revenue above USD 20 million. The full-year 2025 revenue was approximately USD 150 million. As of that same month: 60 million registered creators, 600 million videos generated, 30,000 enterprise users, API services to 10,000 corporate clients worldwide.

Kuaishou management has publicly guided for 100%+ year-over-year revenue growth in 2026. Kling was showcased at CES 2026 — a notable move for a Chinese AI company asserting global positioning — and Kling 3.0 shipped one month later.

These figures place Kling in a different commercial category than its open-source video contemporaries (HunyuanVideo, Wan, CogVideoX) and in the same league as Runway in terms of commercial validation. The iteration cadence — a major version roughly every 1–3 months since launch — is also consistent with a well-resourced engineering organization, not a research lab releasing annual checkpoints.

Limitations and Honest Caveats

Closed-source: There are no open weights, no inference code, no fine-tuning support. Everything runs through Kuaishou’s infrastructure. Users who need to run models locally, fine-tune on proprietary data, or deploy on their own infrastructure have no path to do that with Kling.

Content filtering: The CAC-mandated content restrictions are not configurable. They affect a narrow but specific set of use cases.

API economics at scale: For high-volume production workflows, the official API commitment structure (~$4,200 upfront) and per-second pricing (~$0.10/second Pro equivalent) make Kling more expensive than self-hosted open-source alternatives at scale. The math changes for teams that don’t have the GPU infrastructure to run HunyuanVideo or Wan locally.

Self-reported benchmarks: The OmniVideo-1.0 benchmark in the Kling-Omni paper is Kuaishou’s own evaluation. Claims of superiority over Veo or Runway should be read as marketing context rather than third-party validation. Independent community evaluations are more reliable for relative quality assessments.

Benchmark scarcity for the 1.x/2.x era: Kuaishou did not publish arXiv papers for the original Kling architecture. The most detailed technical disclosure came with the Kling-Omni report in December 2025. For the first eighteen months, understanding what made Kling’s physics simulation work required inference from product demonstrations rather than published methods.

Free tier privacy: Public queue, watermarked output, and visible prompts on the free tier make it unsuitable for professional work involving real people or confidential material.

Rating: 4/5

Kling earns a 4 out of 5.

The case for a high rating: Physics-coherent motion (fluid, cloth, hair) that consistently benchmarks at or near the top of commercial systems; Motion Brush trajectory control that enables a qualitatively different level of creative direction; an iteration cadence that has delivered native 4K, multi-shot storyboarding, and synchronized audio within nineteen months of launch; the most commercially validated AI video platform outside the US; and a version history that suggests this trajectory continues.

The case against a higher rating: Closed-source with no path to local deployment or fine-tuning; content filtering on politically sensitive topics; API pricing that disadvantages high-volume production users; and the largest technical contribution — the Kling-Omni architecture — was not publicly documented for the first eighteen months of the product’s existence. At the 4/5 level, the limitations are real but do not change the fundamental assessment that Kling is a capable, commercially validated, actively developed professional tool.

For teams building at the scale of the open-source ecosystem or needing to deploy models on their own infrastructure, HunyuanVideo or Wan 2.1 are the relevant comparisons and have no content restrictions. For commercial creative production teams where per-clip cost and convenience matter more than infrastructure flexibility, Kling is among the most capable tools available.

ChatForest reviews AI tools based on public documentation, technical reports, developer community sources, and press coverage. We do not conduct hands-on testing. Review reflects information available as of May 2026.