Kuaishou’s Kling 4.0, which became generally available on 15 October 2026, and MiniMax’s Hailuo 2.3, which arrived in April 2026, represent a six-month gap that has amplified rather than resolved the differences between two fundamentally different approaches to AI video generation. According to OrcaRouter, the newer model does not simply replace the older one, because the two were built for different jobs and the passage of time has made that distinction sharper.
Key takeaways
- Kling 4.0 generates up to 30-second clips with stereo sound and ten keyframes, launched 15 October 2026
- Hailuo 2.3 delivers native 1080p stylised video at 6 seconds (1080p) or 10 seconds (768p) but includes no audio
- MiniMax positions Hailuo 2.3 explicitly for cinematic camera motion, human micro-expressions and anime-adjacent content, not photorealism
- Kling 4.0 supports dialogue-synced clips suitable for direct publishing; Hailuo 2.3 requires downstream audio addition
- The six-month development window has widened use-case divergence rather than creating a simple upgrade path
Two Models, Two Workflows
The comparison that writes itself—newer model wins—misses the architecture of the decision, as OrcaRouter observes. Hailuo 2.3 is a stylist with native 1080p output and no audio generation. Kling 4.0 is a controller offering 30-second takes, two-channel stereo sound and ten-keyframe sequencing. If the deliverable is anime-adjacent motion work that will receive music and dialogue in post-production, the older model remains the better instrument for that specific workflow. If the requirement is a twenty-second take with synchronised dialogue, Hailuo 2.3 cannot fulfill the brief at any price point.
As OrcaRouter notes, the useful question is not which model is objectively superior but which limitation aligns with the production pipeline in question. Both models present constraints; the decision rests on which constraint disrupts the workflow least.
What Hailuo 2.3 Delivers
MiniMax’s model is explicit about its target: cinematic camera motion, human motion and micro-expressions, and stylised or anime content are the design strengths, whilst photorealism is not the objective. That represents a substantial market segment, given that a significant share of short-form video content is stylised rather than photorealistic, and on that ground Hailuo 2.3 competes with AI models that cost several times more per second of output.
The specification reveals the model’s April 2026 vintage. Resolution is native 1080p—MiniMax does not attach an upscaling caveat, unlike several competitors—but clip length is constrained: 6 seconds at 1080p, 10 seconds at 768p. There is no end-frame control, which makes multi-shot continuity a manual stitching operation. The model includes no native audio generation; an earlier Hailuo release supported audio, but version 2.3 removed the feature, so every generated clip requires sound to be added in a separate editing pass.
What Kling 4.0 Brings
Kling 4.0, from Kuaishou, generates both picture and two-channel stereo sound in a single inference pass. The model holds a continuous take for up to 30 seconds and supports ten-keyframe sequencing, which allows multi-stage direction within one clip. According to OrcaRouter, the design expectation is that output is publication-ready rather than an intermediate plate for finishing.
That architectural choice shapes every downstream decision. A clip with synchronised dialogue, ambient sound and continuous 30-second narrative continuity can go directly into a social feed or presentation without visiting an audio editor. The trade-off is that Kling 4.0’s styling strengths differ from Hailuo’s: where MiniMax explicitly targets stylised and anime content, Kuaishou’s model is optimised for scenarios where audio-visual synchronisation is non-negotiable.
As reported by Intelligent Living, Kling 4.0 has been deployed for sustainability-focused video content that requires voiceover or natural soundscapes without the “lecture” tone that text-heavy explainer videos can carry. The 30-second duration and native audio capability make the model suitable for narrative-driven environmental storytelling where pacing and sound design are load-bearing elements of the message.
Benchmark Visibility and the Scoring Gap
Neither model can be compared on an independent third-party score today, and for different reasons. As OrcaRouter reports, Luma Ray 3.2—a model in the same general release window—has no row on any of Artificial Analysis’s three video leaderboards (text-to-video, image-to-video or editing) as of 22 October 2026. Kling 4.0 was released eight days prior to that check and has not yet been entered into the independent benchmark cycle.
Hailuo 2.3, released in April, also does not appear on the current leaderboards. The absence of standardised scoring means that cost and capability comparisons rely on specification sheets and workflow testing rather than Elo ratings or head-to-head preference studies. For organisations evaluating AI API costs for video generation at scale, the lack of benchmark parity means pilot testing against real production prompts becomes the primary decision input.
Specification Comparison
| Feature | Kling 4.0 | Hailuo 2.3 |
|---|---|---|
| Release date | 15 October 2026 | April 2026 |
| Maximum clip length | 30 seconds | 6 seconds (1080p) / 10 seconds (768p) |
| Audio generation | Stereo (two-channel) | None (removed from earlier versions) |
| Native resolution | Not specified in sources | 1080p (native, no upscale caveat) |
| Keyframe control | Ten keyframes | No end-frame control |
| Design focus | Audio-visual sync, longer narratives | Cinematic motion, stylised/anime, micro-expressions |
| Photorealism target | Not specified in sources | Explicitly not the objective (MiniMax) |
Use-Case Fit Over Feature Parity
The six-month gap between Hailuo 2.3 and Kling 4.0 has not created a linear progression but rather a branching divergence. If the production workflow involves generating stylised motion sequences—anime-style character animation, cinematic camera movements, exaggerated human micro-expressions—and adding a separately licensed music track or voiceover in post, Hailuo 2.3 delivers native 1080p output optimised for that pipeline. The absence of audio is not a defect in that context; it is the expected workflow.
If the deliverable is a self-contained clip with dialogue, ambient sound and 20- to 30-second narrative continuity—corporate explainer videos, social media storytelling, environmental awareness content of the kind Intelligent Living describes—then Kling 4.0’s integrated audio and extended duration are not convenience features but table stakes. Hailuo 2.3 cannot generate synchronised dialogue, so it cannot complete that job regardless of its strengths in other dimensions.
For teams evaluating which model to integrate into production, the decision tree begins not with release date or vendor but with the question of whether audio is generated or added. That single architectural choice cascades through every subsequent consideration: clip length, keyframe control, styling focus and post-production tooling. Organisations comparing AI price-performance across video generation models will find that cost per second is less informative than cost per finished deliverable, which depends entirely on how much downstream editing each model requires.
Frequently Asked Questions
What is the main difference between Kling 4.0 and Hailuo 2.3? Kling 4.0 generates up to 30 seconds of video with stereo audio in one pass, whilst Hailuo 2.3 produces 6- to 10-second clips at native 1080p without audio, optimised for stylised and anime-adjacent content that will receive sound in post-production.
Which model was released more recently? Kling 4.0 became generally available on 15 October 2026, six months after Hailuo 2.3 arrived in April 2026.
Can Hailuo 2.3 generate audio? No. According to OrcaRouter, Hailuo 2.3 does not include native audio generation; an earlier Hailuo version supported audio, but MiniMax removed the feature in version 2.3.
Which model is better for anime-style content? MiniMax explicitly positions Hailuo 2.3 for stylised and anime content, cinematic camera motion and human micro-expressions, making it the design-fit choice for that use case where audio will be added separately.
How long can each model’s clips run? Kling 4.0 supports continuous takes up to 30 seconds with ten-keyframe sequencing. Hailuo 2.3 generates 6 seconds at 1080p or 10 seconds at 768p with no end-frame control for multi-shot continuity.
The Bottom Line
Kling 4.0 and Hailuo 2.3 do not compete on a single performance axis. One generates audio and extends to 30 seconds; the other delivers native 1080p stylised motion in shorter takes and expects downstream audio work. The six months between their releases have widened the workflow divergence rather than producing a superseding upgrade. Teams selecting between them are not choosing the newer model over the older but matching architectural assumptions—audio integration, clip length, styling focus—to the production pipeline that will receive the output. Both models present trade-offs; neither is universally preferable. The decision rests on which limitation the workflow can accommodate and which capability it cannot function without.
Sources: intelligentliving.co, orcarouter.ai. Reported 22 October 2026.