Kandinsky 6.0 Video With Synced Audio
Synced Audio

Kandinsky 6.0 Video is a family of open-source diffusion models from Kandinsky Lab that generate video together with a synchronized soundtrack. The lineup has two members: Kandinsky 6.0 Video Pro at 29 billion parameters and Kandinsky 6.0 Video Lite at 3 billion. Both output 5-second clips with 44 kHz audio including lip-sync, in text-to-audio-video and image-to-audio-video modes, and both ship with a super-resolution stage that raises output to Full HD.

Yes. The code and the model weights are released under the MIT license, which is one of the most permissive licenses in common use — commercial products, fine-tunes and redistribution are all allowed. Free to license is not the same as free to run, though: you still pay for the GPU time, whether that is hardware you own or accelerators you rent by the hour.

It depends entirely on which checkpoint you choose. The 3B Lite model is the one aimed at modest setups and is the practical choice for a single high-end consumer GPU. The 29B Pro model is a different class of workload and realistically wants professional datacenter accelerators with substantially more VRAM. If you are unsure, start with Lite and measure before committing to Pro-scale hardware.

Each generation is a 5-second clip with its own synchronized audio. That is a shot, not a scene, so longer pieces are built by generating several clips and cutting them together in an editor. Planning in 5-second beats from the start — rather than trying to force one continuous take — produces noticeably better results.

No — we want to be straight about that. VO3 does not host Kandinsky 6.0 Video. VO3 runs its own lineup of hosted models that also produce native synchronized audio, including Sora 2, VEO 3.1 and WAN 2.5 Audio. If you specifically need Kandinsky 6.0 Video weights, get them from Hugging Face and run them yourself; if what you actually need is a finished clip with sound in a browser, VO3 will get you there faster.

Choose Kandinsky 6.0 Video when control is the requirement: data that cannot leave your network, fine-tuning on proprietary footage, reproducible research, or per-clip costs you want to drive down at high volume once the hardware is paid for. Choose a hosted model when throughput and deadlines are the requirement, since there is no environment to build, no driver to debug and no queue of your own to manage. Plenty of teams do both — open weights for experiments, hosted for client deliverables.

Conventional pipelines generate a silent video and then add sound in a second tool, which means every footstep, consonant and musical hit has to be manually aligned. Because Kandinsky 6.0 Video generates both streams jointly through its dual-stream CrossDiT architecture, the audio arrives already aligned to the picture and character lips match the dialogue. For dialogue-driven work this is the difference between minutes of review and hours of sound editing.

The weights are published on Hugging Face under the kandinskylab organization, including Diffusers-format variants of the Lite model, and the inference code is on GitHub. There is also a technical paper describing the architecture and training approach if you want the details before you commit hardware to it.