Kling 4.0 First Look: What Kuaishou's New Video Model Actually Does

Kuaishou's Kling 4.0 Flash video model is in early access for Kling Ultra annual subscribers, with the full model expected in October, and hands-on testing shows lip sync drift that plagued Kling 3.0 is largely fixed. Testers found the Omni reference system now supports up to 10 key frame references per generation, spatial consistency improved modestly, and resolution is capped at 720p in the early-access build. Kling 4.0 follows Kling 3.0, which launched in February, and a Turbo variant released in June.

Kling 4.0 First Look: What Kuaishou's New Video Model Actually Does Kling 4.0 Flash is in early access. Here's what hands-on testing shows about its text-to-video, image-to-video, lip sync, and Omni reference tools. What is Kling 4.0? Kling 4.0 is the newest generation of Kuaishou’s AI video model, currently rolling out in early access as Kling 4.0 Flash. It follows Kling 3.0, which launched in February, and a Turbo variant released in June. Right now, Flash access is limited to Kling Ultra subscribers on the annual plan, with a wider release of the full model expected in October. Early hands-on testing of the Flash version covers text-to-video, image-to-video, lip sync, and Kling’s Omni reference system, giving a preview of where the model stands before general availability. TL;DR - Kling 4.0 Flash is the early-access version of Kuaishou’s new video model, available now only to Ultra annual subscribers, with the full model due in October. - Text-to-video quality looks solid in early tests, though generations can introduce small unprompted actions the model invents on its own. - Image-to-video performance held up well on a dialogue-heavy test clip, including inventing a consistent appearance for a character whose face wasn’t visible in the source image. - Lip sync drift , a persistent problem in Kling 3.0, appears to be largely fixed in 4.0 based on dialogue-driven test generations. - Omni reference mode now supports up to 10 key frame references per generation, and combining key frames with Omni references seems to improve character performance consistency. - Spatial awareness showed modest improvement, with object placement staying consistent across different points in the same generated scene. - Resolution is capped at 720p for now in the early access build, which is a limitation worth noting before the full model ships. Other agents start typing. Remy starts asking. Scoping, trade-offs, edge cases — the real work. Before a line of code. What’s new in Kling 4.0 compared to earlier versions? The headline improvement in early testing is lip sync. Kling 3.0 had a well-known issue where lip sync would drift out of alignment over the course of a longer clip, especially with extended dialogue. In testing against a dialogue-heavy scene an FBI agent having an extended conversation in a diner , Kling 4.0 handled the full exchange without the sync breaking down, suggesting Kuaishou made real progress on this front. Spatial consistency also appears improved. In one test, a character walks into a room, interacts with objects, and later returns to the same space. A kitchen sink referenced early in the generation stayed in the same position later in the clip, a small but meaningful sign that the model is tracking scene geometry more reliably than before. The Omni reference system, Kling’s tool for maintaining consistent characters across generations, now supports up to 10 key frame references in a single generation. Testers found that combining key frames with Omni references produced better results than using Omni reference alone, apparently giving the model a stronger sense of how a character should move and perform, not just what they should look like. How does Kling 4.0 handle text-to-video? In basic text-to-video tests, such as a scene depicting a man in a suit being stopped for jaywalking, Kling 4.0 produced generally solid results. One recurring quirk: the model sometimes adds small actions or business that weren’t in the prompt at all, like a police officer making a gesture that looked like he was trying to identify someone from a photo. These unprompted details aren’t necessarily wrong, but they’re a reminder that the model is filling in narrative gaps on its own, which can be a feature or a bug depending on what you’re trying to produce. How good is Kling 4.0 at image-to-video? Image-to-video testing focused on a classic benchmark case: a scene with dialogue where a character’s back is turned to the camera in the source image, forcing the model to invent their face and expression once they turn around. Kling 4.0 handled this well, maintaining stylistic consistency throughout the generated clip while also managing a full exchange of dialogue without breaking down. That said, testers caution that some decoherence visual drift or inconsistency is still likely in longer or more complex generations, particularly since this is the Flash variant rather than the full model. Flash models generally trade some fidelity for speed and lower cost, so results from the full Kling 4.0 release later this year may differ. What is Kling’s Omni reference mode and how does it perform? Omni reference is Kling’s system for keeping a character visually consistent across different generations and scenes, functioning similarly to a character reference sheet the model can draw from repeatedly. Each major video model has its own take on this kind of feature, and Kling’s version has consistently been a strong performer for character consistency. Remy doesn't write the code. It manages the agents who do. Remy runs the project. The specialists do the work. You work with the PM, not the implementers. In testing, an established character “flamethrower girl” was carried into Kling 4.0’s Omni mode with mixed but promising results. Minor physics details were off a candle that should have melted during a fire effect didn’t , but the core character consistency and spatial logic held up across the scene. A separate noir-style detective scene highlighted both the strengths and quirks of Omni mode. Eyelines matched correctly across cuts, which matters a lot for believability in dialogue scenes, but the model also added a gray jacket to a character that wasn’t part of the original reference image or prompt. It’s a small inconsistency, but a useful reminder that Omni reference systems still improvise when they don’t have explicit visual direction. Video quality in these Omni tests capped out at 720p in the current early access build, and some testers noted that character performances could feel a bit flat emotionally, even when the technical elements lip sync, eyelines, continuity were working correctly. Is Kling 4.0 worth using right now? For most people, not yet, simply because access is restricted to Ultra annual subscribers during this early access phase. For those who do have access, Kling 4.0 Flash represents a solid incremental step forward, particularly on lip sync accuracy and spatial consistency, two areas that have historically been weak points for AI video models generally, not just Kling. The bigger test will come when the full non-Flash Kling 4.0 model releases, expected around October. Flash models are built for speed and lower compute cost, so they typically sacrifice some visual fidelity and complexity compared to a flagship release. Early impressions suggest Kling 4.0 is a meaningful upgrade over 3.0, but not necessarily a dramatic leap over the best competing video models currently available. Anyone evaluating it for production work should wait for the full model and broader access before drawing firm conclusions. Frequently Asked Questions When does Kling 4.0 fully release? The full Kling 4.0 model, as opposed to the current Flash early access version, is expected to roll out sometime in October, based on Kuaishou’s stated timeline. What’s the difference between Kling 4.0 and Kling 4.0 Flash? Flash is a faster, lighter version of the model typically optimized for speed and cost rather than maximum quality. It’s common for Kling to release a Flash variant to early access subscribers before the full flagship model becomes widely available. Does Kling 4.0 fix the lip sync problems from Kling 3.0? Early testing suggests yes. Kling 3.0 had a known issue with lip sync drifting out of alignment during longer dialogue sequences, and dialogue-heavy test clips in Kling 4.0 Flash did not show the same breakdown. What is Omni reference in Kling? Omni reference is Kling’s system for keeping a character’s appearance consistent across different video generations, functioning like a reusable character reference. Kling 4.0 extends this to support up to 10 key frame references per generation. What resolution does Kling 4.0 currently support? In the current early access Flash build, video output is capped at 720p. It’s unclear whether the full model release in October will support higher resolutions.