Kandinsky 6.0 Video: Open models generate video with matching sound

What's going on

In early October 2026, Kandinsky Lab released Kandinsky 6.0 Video, a model family that generates video and audio together. There are two variants: Kandinsky 6.0 Video Pro with 29 billion parameters (parameters are, roughly speaking, the adjustable values a model learns during training) and Kandinsky 6.0 Video Lite with 3 billion parameters. Code and model weights are released under the MIT license. That means anyone can use them for free and build their own products on top. The weights are on Hugging Face, the code on GitHub.

Both models generate videos with dialogue, music and ambient sound. The audio matches what happens on screen, and the characters' lip movements match the speech. You can describe a scene in text, or upload a first frame and describe the sound you want. You can also generate video without audio. The maximum length of a clip with audio is currently five seconds. The developers plan to extend that in later versions.

What the models can do

According to the sources, the audio covers a wide range: speech, sound effects, footsteps, wind or passing cars. The audio output has a sampling rate of 44 kHz, the standard for many streaming services. According to the report, perceived quality depends on the prompt and the generation conditions.

The developers list these use cases:

• Animating family photos
• Video greetings with speech or music
• ASMR scenes, such as rustling paper or crackling firewood
• Bringing existing images or film scenes to life

The models are also meant to render the physical motion of people, animals and objects more naturally, especially during fast movement or camera cuts. That is a claim by the maker.

Pro vs. Lite

Base generation runs at SD resolution. A built-in super-resolution module then upscales the image to HD or Full HD. That lets you trade speed against detail. Lite is aimed at NVIDIA consumer graphics cards, with presets for 16, 24 and 32 GB of video memory. With block offloading, both models can also be started on a home PC. Parts of the weights are passed through the GPU one after another. How fast it runs and how much memory it needs depends on your configuration.

Architecture and training

Kandinsky 6.0 Video consists of two connected modules for video and audio. According to a third-party source, it is a dual-stream CrossDiT: a pretrained video stream is linked to a newly trained audio stream via bidirectional cross-attention. Cross-attention is a mechanism that lets two data streams "pay attention" to each other. That way, image and sound are generated together rather than built separately and stitched together afterward.

The audio module was first trained from scratch on 40 million audio tracks. Joint training with the video module followed, on 7 million paired audio-video segments. The team picked clips with clear audio and close-ups of people speaking, to better align speech and facial expressions. According to the third-party source, fine-tuning, reinforcement learning (learning through rewards) and distillation (shrinking or speeding up a model) followed. Reinforcement learning cut the word error rate of the generated speech by 47% in the Pro model. That figure comes from the team itself. For the metric definition and test conditions, the article points to the technical report.

Comparisons and benchmarks

The technical report includes comparisons made by human raters. According to it, Pro beats Kandinsky 5.0 Video Pro on every criterion. Gearbrain cites further results said to come from the report:

• Against the open LTX 2.5, Pro leads on image quality and speech quality.
• Among open models, only MiniMax H3 is ahead.
• In image animation, Pro reportedly beats the closed Veo 3.1 Fast on image quality, fidelity to the source image and camera control.

The Digital Journal article urges caution: results vary by task and criterion and are not a general ranking. A third-party source notes that Kandinsky 6.0 Video doesn't appear on the independent Physics-IQ board and that the claims rest on a comparison run by the maker. According to the developers, the previous version led the open models in the text-to-video category on Arena.ai for more than six months. On its own, that says nothing about other tasks.

What you need to try it

• Graphics card: For Lite, an NVIDIA card with 16, 24 or 32 GB of VRAM, such as the RTX 5060 Ti through the RTX 5090. Pro (29B parameters) needs much more: data center cards or heavy offloading.
• Software: A third-party source lists Python 3.13 or 3.14 and an NVIDIA GPU.
• RAM and disk space: The sources give no exact figures. With block offloading you need more system RAM, and generation takes longer.
• Without your own hardware: Pro is available through the FAL platform, and there is a Hugging Face Space for the distilled Pro model. Mac or CPU-only operation is not mentioned.
• Integration: Diffusers, ComfyUI, FastVideo, SGLang and vLLM-omni.

Our take

Open video models with built-in, synchronized audio are rare. Support for five inference stacks on day one lowers the barrier to entry considerably. The MIT license also allows commercial use, fine-tuning and quantization. That is a clear advantage over closed services.

The short five-second clip length and the lack of independent verification are drawbacks. Almost all quality claims come from the developers' own report and rest on human comparisons under their test setup. You'll have to check for yourself whether results on your own hardware look the same. The Lite model is especially interesting because it brings video with sound to home users.