Kling 4.0 Officially Launches: 10-bit HDR and 30-Second Native Output Reset Video Model Benchmarks

The video generation race has hit another key milestone. Kuaishou's Kling AI has officially launched its next-generation video generation model, Kling 4.0, focusing its upgrades on high-dynamic, motion-heavy creative scenarios and making concentrated breakthroughs around three directions: "realism, controllability, and professionalism."

High Dynamics and Complex Camera Movement: From "One Stunning Frame" to "Stable Delivery"

Over the past year, the most frequently cited shortcoming of AI video generation hasn't been image quality but stability—high-speed motion tends to blur, complex camera movements tend to collapse, and multi-character interactions struggle to stay consistent. Kling 4.0 aims its main firepower squarely at these high-frequency pain points.

According to official descriptions, the new model delivers more stable performance on large-scale dynamics, with improved coherence in high-speed motion and continuous action, while tracking shots, orbiting shots, and other large camera movements maintain image stability. Kling AI emphasizes in its official statement that professional creators need "not just the occasional stunning image, but the ability to reliably nail the key shots that determine a work's quality," and to make generated footage usable within a complete production pipeline. This positioning is crucial: it signals a shift in product goals from "generation tool" to "production-grade delivery."

In terms of audiovisual quality, Kling 4.0 supports up to 10-bit HDR output, delivering richer color gradation and smoother transitions between light and shadow. In high-contrast scenes such as backlighting, night shots, and neon lighting, it preserves more detail in both highlights and shadows, leaving more room for post-production color grading. It also supports 4K, 1080p, and 21:9 ultra-wide aspect ratios, further expanding compositional possibilities.

10-bit HDR refers to HDR output with 10-bit color depth. Traditional 8-bit video has only 256 gradations per color channel, which easily produces banding and posterization in high-contrast or large-gradient scenes like skies, neon, and night shots; 10-bit boosts each channel to 1024 levels, capturing roughly 64 times the total color count of 8-bit. Combined with HDR's high dynamic range, it preserves more detail in both highlights and shadows simultaneously. For professional production pipelines, this means Log or HDR footage has greater latitude during color grading, and keying, compositing, and secondary color correction won't suffer from posterization due to insufficient data—one of the technical prerequisites for "generated footage to be usable within a complete production pipeline."

30-Second Native Output and Director-Level Thinking: The Possibilities of Long-Take Storytelling

Extending single-pass native generation to a maximum of 30 seconds is the most underappreciated upgrade in this release. Duration itself isn't the goal, but it directly determines the minimum unit of storytelling—10 seconds can only capture "a moment," while 30 seconds offers the chance to complete an action sequence, a scene change, or an emotional arc.

At the shot choreography level, Kling 4.0 enhances understanding of push-in, pull-out, pan, tracking, follow, and orbit movements, with smoother and more natural transitions between shot sizes and scenes. The model also introduces director-level thinking to reasonably expand and supplement prompts, with text input length increased to a maximum of 8000 tokens, enabling fuller and more precise communication of creative intent. The practical significance: the creator's job is shifting from "writing clear parameters" to "clearly stating what you want to shoot."

Token is the basic unit of text processed by large models; one token roughly corresponds to 0.75 English words or 1–2 Chinese characters, so 8000 tokens amounts to a prompt capacity of several thousand Chinese characters. For video generation, prompt length directly determines "how complex a choreography can be described": shot types, motion trajectories, shot size changes, character action timing, lighting and atmosphere all rely on text constraints. When capacity is insufficient, the model can only fill in the gaps itself, leading to semantic drift. After reading a long prompt, Kling uses "director-level thinking" to expand and structurally complete it—essentially rewriting verbose natural language into storyboard-style structured instructions, keeping the creator's intent coherent throughout the entire video.

All-Around Reference: 15-Item Material Combinations and 10 Keyframes

Controllability is essential for professional creation, and Kling 4.0 has significantly expanded in this area.

The newly upgraded all-around reference feature supports flexible combinations of images, videos, subjects, and voice profiles in a single task, up to 15 items total: up to 10 images, 5 videos, and 7 subjects. It supports various image formats including reference subject images, three-view drawings, grid images, storyboards, and wireframes, as well as video input referencing character performance, actions, camera movement, and narrative rhythm. White-model videos and depth-map videos can also be passed in as reference materials. Multiple reference dimensions work in coordination, making characters, actions, camera work, and sound closer to expectations.

Creative replication has also been upgraded, allowing creators to draw on the narrative rhythm, shot arrangement, and expressive approach of reference videos, and replace the products, characters, and materials within them—useful for commercial video creation in product ads, social media content, fashion and beauty, and consumer electronics. Multi-keyframe precise control supports up to 10 keyframe inputs, letting creators freely set key image nodes and specify character states, scene changes, and plot points. Video continuation will subsequently support multiple rounds, extending up to 2 minutes.

From Film Festivals to Ad Festivals: Cross-Validation in Professional Scenarios

This generation's capabilities aren't just described at the parameter level—they've already been used in multiple professional scenarios.

At the 31st Busan International Film Festival's Asian Contents & Film Market (ACFM), director Kim Seong-su of 12.12: The Day shared Kling AI's application in actual production. In his in-progress AI short film Worm Train, there's a core shot of passengers crammed into the same carriage as a giant insect, bodies constantly colliding. After Kling 4.0 entered internal testing, Director Kim and his team completed the presentation of this scene. He stated: "With the upgrades in Kling 4.0, we can get closer to true cinematic creation, and AI-generated imagery now possesses richer cinematic expressiveness."

At the Toronto International Film Festival market, the in-production MINIBOTS integrated Kling 4.0 into its professional production pipeline, using features like all-around reference to complete action design and shot choreography. Evolutionary Films co-founders John Adams and Diane Shorthouse commented: "The consistency between camera movement and multiple characters is especially impressive—it feels more and more like directing a film rather than just writing prompts for AI."

There are also real-world cases in advertising. This summer's Cannes Lions International Festival of Creativity included AI in its evaluation system for the first time, and two works produced with Kling AI won Cannes Gold Lions.

MCP and CLI Integration, API Opening Soon

One easily overlooked but far-reaching change in this upgrade: Kling 4.0 has integrated with Kling MCP and CLI, allowing creators to invoke 4.0's all-around reference, keyframe-to-video, and other capabilities through AI Agents. At the same time, Kling 4.0's model capabilities will soon open to enterprise customers and developers through an API.

This means video generation capabilities are being packaged into programmable modules embedded in more complex workflows. For developers, video generation is no longer a standalone web operation but a link that can be orchestrated by agents and arranged within pipelines.

According to public information, Kling AI has completed over 30 major iterations since its inception, attracting more than 100 million users across 224 countries and regions worldwide, serving over 50,000 enterprises, with annual revenue reaching $500 million. Competition in video AI is shifting from "can it generate" to "can it generate stably and controllably," and Kling 4.0's upgrade is another footnote to this shift.

MCP (Model Context Protocol) is an open protocol that lets AI Agents invoke external tools and data sources in a unified way; CLI (Command-Line Interface) allows developers to trigger model capabilities directly through scripts or terminals. Together, they mean Kling 4.0's capabilities are no longer limited to interactive web-based operations but can be packaged as tools, autonomously orchestrated by Agents, and batch-executed by scripts. Typical uses include: an Agent reading a storyboard and automatically generating shot by shot, batch-replacing product materials for multi-version ad testing, and plugging video generation nodes into existing post-production pipelines. The accompanying API opening lets enterprises embed video generation into their own systems, transforming it from "a web tool" into "a programmable production pipeline."

One-line Summary

Kling AI launches Kling 4.0, focusing on high-dynamic stability, 30-second long takes, and all-around reference controllability, with MCP and CLI integration.

Paragraph Summary

Kling AI's next-generation video generation model, Kling 4.0, is officially online, upgraded around "realism, controllability, and professionalism." Addressing high-frequency pain points like blurring in high-speed motion, collapsing in complex camera work, and difficulty maintaining multi-character consistency, the new model improves stability in large-scale dynamics and tracking/orbiting shots, and supports up to 10-bit HDR and 4K/1080p/21:9 formats. Single-pass native generation extends to a maximum of 30 seconds, text input caps at 8000 tokens, and director-level thinking is introduced to expand prompts. On controllability, all-around reference supports up to 15 material combinations per task, multi-keyframe control supports up to 10 keyframes, and continuation extends to 2 minutes. The model has integrated with Kling MCP and CLI and will soon open through an API, with real-world cases at Busan, Toronto film festivals, and Cannes Lions.