Kling 3.0

Overview

Kling 3.0 is Kuaishou's closed-weight video generation family, released on February 5, 2026. The launch press release — "Kling AI Launches 3.0 Model, Ushering in an Era Where Everyone Can Be a Director" — introduces four SKUs at once: Video 3.0, Video 3.0 Omni, Image 3.0 and Image 3.0 Omni. Kuaishou describes the series as supporting "full multimodal input and output spanning text, images, audio, and video," with "extended video duration of up to 15 seconds, and native audio generation across multiple languages, dialects, and accents."

The headline architectural pitch is that generation, editing and audio are one model rather than a pipeline. Video 3.0 Omni takes a reference video and extracts "visual traits and voice characteristics" of a character to carry across new scenes, and exposes a "multi-shot storyboard feature that allows users to generate professional shots where they can specify the duration, shot size, perspective, narrative content and camera movements." Image 3.0 and Image 3.0 Omni add 2K and 4K image generation. At launch the models were gated to Ultra subscribers before a public rollout.

The strongest part of the Kling story is not the leaderboard, it is the P&L. Kuaishou's first-quarter 2026 results release reports "Kling AI generating over RMB650 million in revenue in the first quarter, representing year-over-year growth of more than 300.0%," and "In March 2026, the annualized revenue run rate (ARR) of Kling AI was approximately USD500 million." The same release credits 3.0 directly: "In February 2026, we launched the Kling AI 3.0 series, extending Kling AI's global leadership in model capabilities and product experience" — that last phrase is a vendor claim, and the benchmarks below do not support it.

Separately, the South China Morning Post reported on July 1, 2026, citing unnamed sources, that Kling AI was nearing a US$3 billion fundraising round at a US$18 billion post-investment valuation, with a Hong Kong listing process to begin "in the next 12 months." The same report names Tencent among the backers. Kuaishou has not confirmed the round in a filing this page could verify, so treat it as reported, not settled.

A note on naming

Three naming traps surround this model.

1. There is no Kling 3.1. Kuaishou has announced Video 3.0, Video 3.0 Omni, Image 3.0, Image 3.0 Omni and — later — Kling 3.0 Turbo. Nothing numbered 3.1. Even kling31api.org, an unaffiliated site built on the name, concedes: "There is no Kling 3.1 release."
2. kling-o3 is not a fourth-generation model. The "O" stands for Omni. kling-v3 is Video 3.0; kling-o3 is Video 3.0 Omni. Both are the 3.0 generation.
3. arena.ai's kling-v3-pro and Artificial Analysis's "Kling 3.0 1080p (Pro)" are the same model on two incompatible Elo scales. Their numbers must never be compared to each other.

Capabilities

• Native audio in the same pass: Dialogue, sound effects and ambient sound are generated with the video, not composited afterwards. Kuaishou lists English, Chinese, Japanese, Korean and Spanish, plus English accents and Chinese dialects, with different languages assignable to different characters in one scene.
• Multi-shot storyboarding: A single prompt can specify per-shot duration, shot size, perspective, narrative content and camera movement, with the model maintaining continuity between shots.
• Reference-based consistency (Omni): Up to seven reference images (four when a reference video is also supplied), or a reference video, to lock a character's appearance and voice across generations. Replicate's listing describes it as "character consistency, multi-shot control, video editing, and native audio in one model."
• Video editing from text: The Omni endpoint accepts an existing video as a base and edits it under natural-language instruction — a capability outside the scope of most text-to-video models.
• Text preservation: The launch release specifically claims improved rendering of legible text in frame, aimed at branded assets.
• Native 4K video: Announced April 29, 2026, after the February launch. Kling's claim, quoted in Korean press: "This is the industry's first model to directly output 4K video at the generation stage."
• Motion transfer: A separate kling-v3-motion-control endpoint transfers character motion from a reference video onto a still image.

Technical Specifications

• API model names: kling-v3 (Video 3.0), kling-o3 (Video 3.0 Omni), plus kling-v3-turbo and kling-v3-motion-control
• Modalities in: Text, images (up to 7 references), reference video (fal's O3 schema documents a 10-second maximum), audio
• Modalities out: Video with synchronized audio; images at 2K/4K
• Duration: 3–15 seconds per generation, 5s default
• Resolution tiers: Standard = 720p, Pro = 1080p, plus a separate native-4K image-to-video endpoint
• Aspect ratios: 16:9, 9:16, 1:1 for text-to-video; determined by the start frame for image-to-video
• Audio languages: English, Chinese, Japanese, Korean, Spanish (+ accents and dialects)
• License: Proprietary. No published weights, no Hugging Face repository
• Closest technical disclosure: the Kling Team's "Kling-Omni Technical Report" (arXiv:2512.16776, December 18, 2025), which describes "a generalist generative framework designed to synthesize high-fidelity videos directly from multimodal visual language inputs"

Kuaishou does not publish a parameter count, a training-compute figure, a frame rate, or a training-data description for Kling 3.0, so this page does not state one. English Wikipedia describes the original 2024 Kling as a diffusion transformer with a 3D variational autoencoder; that description has not been re-confirmed for the 3.0 generation.

Use Cases

• Short-form social video: 9:16 clips up to 15 seconds with spoken dialogue and lip sync generated in one call — the workload that put Kling at "the number one position on the App Store across 42 countries and regions," per Kuaishou's Q1 2026 results release.
• Advertising and product motion: Image-to-video from a product still, with text preservation for logos and packaging copy.
• Previsualization and storyboarding: The multi-shot mode produces a rough cut of a scene with specified shot sizes and camera moves before any production spend.
• Character-consistent series content: Omni's reference-video conditioning keeps a recurring character's face and voice stable across episodes.
• Localized dubbing at generation time: Multi-character dialogue where each character speaks a different language, generated rather than dubbed.
• Enterprise creative pipelines: Kuaishou claims partnerships with "more than 30,000 enterprise clients," concentrated in film and advertising.

Performance / Benchmarks

Kuaishou publishes no benchmark table for Kling 3.0. The only independent numbers are human-preference Elo scores from two unrelated organizations. Their scales are not comparable — arena.ai (the former LMArena, rebranded January 28, 2026) and Artificial Analysis run different arenas with different populations and different anchoring.

arena.ai — Image-to-Video Arena (data as displayed July 8, 2026)

Kling has the sixth-largest vote count on that board — behind grok-imagine-video-720p (462,748), hailuo-2.3 (229,374), seedance-v1.5-pro (218,351), kling-2.6-pro (182,012) and ltx-2-19b (166,117) — and the 13th-highest score. On arena.ai's text-to-video board (data as displayed July 8, 2026), no Kling 3.0 entry appears at all; the highest-ranked Kling model is kling-2.5-turbo-1080p at #22 with 1219 ±17, followed by kling-2.6-pro at #23 with 1218 ±7.

Artificial Analysis — Video Arena (different Elo scale, do not compare to the above)

On the no-audio text-to-video board, Artificial Analysis puts HappyHorse-1.0 (1290), HappyHorse-1.1 (1279), Dreamina Seedance 2.0 720p (1271) and Wan2.7-260612 (1246) ahead of Kling 3.0 1080p (Pro) at 1244.

Correcting a widely repeated claim

"Kling 3.0 is the #1 text-to-video model with an Elo of 1243" is false. The figure traces back to Artificial Analysis, whose model page for Kling 3.0 1080p (Pro) shows an Arena Elo of 1243.83 — and whose leaderboard puts that score at rank 5, not rank 1, on the no-audio board. On the separate with-audio board the model is also rank 5, at Elo 1109.

A second confusion feeds the same myth: llm-stats.com does show Kling v3 in first place, but with an "arena score" of 2023 derived from TrueSkill (a conservative μ − 3σ rating), not Elo, over just 1,104 blind votes across 11 models (as of July 8, 2026). That is a different metric, a different model pool, and under 0.1% of the 1,350,288 votes behind arena.ai's image-to-video board. It does not license the claim.

Limitations

• Closed weights, no self-hosting. There is no checkpoint, no Hugging Face repo, and no path to running Kling on your own hardware. Every generation is a call to Kuaishou or a distribution partner.
• No technical disclosure for 3.0. No parameter count, no frame rate, no training-data statement, no evaluation methodology. The closest artifact is a December 2025 Kling-Omni report predating the launch.
• Not the leader on either independent board. #13 on arena.ai image-to-video, #5 on both the no-audio and the with-audio Artificial Analysis text-to-video boards, #11/#12 on its image-to-video board. Vendor language about "global leadership in model capabilities" is a Kuaishou claim and is not corroborated.
• Audio is not where it wins. Kling 3.0 1080p (Pro) ranks 5th whether or not raters can hear the generated audio, so the modality Kuaishou markets hardest buys it no rank. (The two boards are scored on separate pools; the 1244 → 1109 score difference is not itself a regression.)
• 15-second ceiling. Longer narratives require stitching, and the multi-shot storyboard divides those 15 seconds rather than extending them.
• Pricing is not publicly machine-readable. Kuaishou's own pricing and API-documentation pages return HTTP 446 to automated clients, and otherwise serve a JavaScript shell with no prices in the HTML, which makes independent price verification depend on resellers.
• Corporate structure in flux. A reported US$3 billion raise and a planned Hong Kong spin-out listing mean pricing, API terms and roadmap sit with an entity whose ownership is actively changing.

Pricing & Access

Kuaishou publishes consumer credit tiers and an API rate card on kling.ai, but that host returns HTTP 446 to automated clients, so those figures could not be verified here and are not restated. The prices below are fal.ai's published USD rates, quoted verbatim from its model pages. fal is a distribution partner, not Kuaishou; rates on Replicate, Krea and the first-party API will differ.

fal's wording for the Pro tier: "For every second of video you generated, you will be charged $0.112 (audio off) or $0.168 (audio on), if voice control is used while generating audio you will be charged $0.196." For 4K: "you will be charged $0.42 regardless of whether audio is on or off."

At 1080p with audio, a 15-second Kling 3.0 Pro clip costs about $2.52 on fal. A 15-second 4K clip costs about $6.30.

Consumer access is through the Kling AI app and web platform on a credit subscription. At launch the 3.0 models were restricted to Ultra subscribers before general release.

Self-hosting: not possible.

Ecosystem & Tools

• Kling AI — the consumer app and web studio (blocks automated clients)
• Kling AI Open Platform API docs — first-party API reference
• fal.ai Kling 3.0 — hosted endpoints for v3, O3, Turbo and 4K, with per-second USD pricing
• Replicate kwaivgi — kling-v3-video (356.5K runs), kling-v3-omni-video (702K runs), kling-v3-motion-control (462.3K runs), as of July 8, 2026
• Kling-Omni Technical Report — the Kling Team's December 18, 2025 paper on the unified generate-and-edit framework, published seven weeks before Kling 3.0
• Kling Team on Hugging Face — Kuaishou's open video work (LivePortrait, SVG-T2I); Kling itself is absent. The older KwaiVGI org now redirects here

Community & Resources

• Kling AI Launches 3.0 Model, Ushering in an Era Where Everyone Can Be a Director — the launch press release, February 5, 2026
• Kuaishou Technology Announces First Quarter 2026 Unaudited Financial Results — the RMB 650M / USD 500M ARR figures
• Kuaishou beats estimates as Kling AI video generator's revenue jumps 300% — SCMP, May 27, 2026
• China's Kling AI nears US$3 billion round at US$18 billion valuation — SCMP, July 2026
• Kling Introduces Native 4K Video Generation Without Upscaling — Seoul Economic Daily, April 29, 2026
• Artificial Analysis Text-to-Video Leaderboard and arena.ai Image-to-Video Leaderboard — the two independent Elo boards
• Compare with Sora 2, Seedance 2.0, HappyHorse 1.1, Vidu Q3 Pro, Wan 2.2, and Hailuo 2.3