Kling Video 2.6 — first Kling AI model with native audio (announcement)

TL;DR

Kling AI’s Day-3 release announces Video 2.6, the first Kling model with native audio generation — joint video+audio in a single model rather than a T2V→V2A pipeline. The launch is paired with a short film “I Have a Secret” co-produced with several Kling Creative Partners (@ViralAiFun, @mad_mask, @JH4TC, @misslaidlaw). No architecture details, parameter counts, or benchmarks are disclosed; the tweet is a product launch with a promo (200 credits for follow+like+RT, 200 Standard-Plan giveaways). The relevance for the team is that Kling 2.6 is now another closed-source datapoint in the joint-audio-video race alongside Veo 3, Sora 2, Seedance 2.0, and Vidu Q3.

Key claims

Kling Video 2.6 is Kling AI’s first model with native audio generation, framed as “See the Sound, Hear the Visual” rather than a chained T2V→V2A pipeline [Tweet body].
The launch demo is a short film “I Have a Secret” co-produced with four Kling Creative Partners [Tweet body].
No technical disclosure — no parameter count, architecture, training data, audio sample rate, max duration, or benchmark numbers are stated [Tweet body].

Method

Not disclosed. The tweet shows a 43-second demo video and lists creative collaborators; there is no accompanying technical report, system card, or model card linked from the tweet. Whether 2.6 uses the dual-stream-DiT-with-bidirectional-cross-attention recipe that the open T2AV cluster has converged on (LTX-2: Efficient Joint Audio-Visual Foundation Model, MOVA: Towards Scalable and Synchronized Video-Audio Generation, SkyReels-V4: Multi-modal Video-Audio Generation, Inpainting and Editing model) or something materially different is exactly the question the Joint audio-video generation concept page already flags as an open one.

Results

No quantitative results in the announcement. The 43-second sample is the only public artifact, and there’s no apples-to-apples benchmark comparison against Veo 3 / Sora 2 / open T2AV. The 290.3K views and ~1.5K reposts are engagement metrics, not capability metrics.

Why it’s interesting

Kling 2.6 is the third closed-source frontier T2AV model to ship in 2025–2026 after Veo 3 and Sora 2, and the Joint audio-video generation concept page already lists Kling 3.0 as one of the closed-source reference points the open cluster (LTX-2, MOVA, SkyReels-V4) is chasing — so this is the canonical “what is the closed-source recipe?” datapoint, even with no architecture revealed. It complements Kling-MotionControl Technical Report (the only previously-filed Kuaishou Kling artifact, on character motion control) by pinning down that the same group is now also shipping joint audio. The lack of any technical detail is itself informative: contrast with Alibaba’s Wan releases (Wan2.2: World's First Open-Source MoE Video Generation Model (launch announcement), Wan-Image: Pushing the Boundaries of Generative Visual Intelligence) where each version ships an arxiv tech report — Kuaishou is staying closer to the OpenAI/Google “product first, paper maybe later” cadence for their flagship line.