Kandinsky 6.0 Video
Kandinsky Lab's 3B and 29B models that make five-second video with synchronized 44 kHz audio and lip-sync. MIT license.
In short. Kandinsky 6.0 Video makes short video clips together with their sound. You describe a scene, or give a first frame, and it returns five seconds of video with matching audio, including speech that follows the lips. Lite is small and fast, Pro is larger and looks better, and a separate upscaler brings clips to Full HD. It is free to use, including commercially, under the MIT license.
Summary
Kandinsky Lab released Kandinsky 6.0 Video on 6 October 2026 under the MIT license, with a technical report, arXiv 2610.05608. It is a family of diffusion models that generate video and sound together. Video Lite has 3 billion parameters and Video Pro has 29 billion, and both make 5-second clips at 24 frames per second with synchronized 44 kHz audio, lip-sync included, from text or from a first-frame image. The base clip is small, 480 by 864 pixels in the examples, and a separate super-resolution model lifts it to Full HD, 1920 by 1080. The native pipeline can also start from a first and a last frame, drive the video from an audio file, or generate audio for a video file, and the lab says the models understand Russian prompts and cultural concepts natively.
Each size comes as a standard checkpoint that runs 50 steps, a distilled one that runs 10, and a pretrain checkpoint for fine-tuning. The lab describes the standard ones as two-stage fine-tuning followed by reinforcement learning. Under the hood, one flow-matching transformer handles video and audio together, with Qwen2.5-VL as the text encoder, which can also expand short prompts, a CLIP text encoder, a HunyuanVideo VAE and an audio VAE with a BigVGAN-style vocoder. Diffusers, ComfyUI, vLLM-Omni, SGLang and FastVideo support it. The Diffusers repositories are large, about 90 GB for Pro and 27.8 GB for Lite, partly because each holds a 16.6 GB text encoder. The README's own timings for a 5-second, 480-by-864 clip with audio are 68 seconds for Lite Distill and 235 seconds for Pro Distill on an RTX 4090, against 955 seconds for the 50-step Pro. A 16 GB RTX 5060 Ti ran Lite with a 4-bit text encoder and Pro with block offloading, at 174 and 718 seconds for the distilled models.
Of the four generation repositories, only the main Pro one is gated behind a short form. The lab says that in side-by-side human evaluation Video Pro clearly outperforms Kandinsky 5.0 Video Pro and stays competitive with leading audio-video models, particularly in speech quality, and we have not checked that. In the 30 days to 11 November 2026, Hugging Face counted about 3,600 downloads across the four generation repositories.
Good for
- Short clips with matching sound and lip-synced speech from a text prompt or a first frame
- Experiments on a single 16 GB GPU with Lite, using the distilled 10-step checkpoint to save time
- Driving a video from an audio file, or adding audio to a video, with the native pipeline
- Research and products that want MIT-licensed weights, Diffusers support and pretrain checkpoints to fine-tune
Watch out for
- Clips are limited to 5 seconds, and the base resolution is small until you add the separate super-resolution model
- Pro is slow on consumer cards, with 955 seconds for one 50-step clip on an RTX 4090 in the README, and the repositories are 80 to 90 GB
- The main Pro repository is gated, although the other three are not, and the native pipeline asks for an NVIDIA GPU
- The lab's comparisons with other audio-video models are its own, and the README publishes no scores in text
- Training data and rights for generated speech and likenesses are not described in the files we read
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Kandinsky 6.0 Video. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Kandinsky 6.0 Video Lite (50 steps) | 3B | 28 GB | 16 GB | Laptop or phone |
| Kandinsky 6.0 Video Lite Distill (10 steps) | 3B | 27 GB | 16 GB | Laptop or phone |
| Kandinsky 6.0 Video Pro (50 steps) | 29B | 90 GB | 103 GB | Big-memory desktop |
| Kandinsky 6.0 Video Pro Distill (10 steps) | 29B | 81 GB | 93 GB | Workstation |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 16 GB of GPU memory.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
Frequently asked questions
- What is Kandinsky 6.0 Video?
- Kandinsky 6.0 Video makes short video clips together with their sound. You describe a scene, or give a first frame, and it returns five seconds of video with matching audio, including speech that follows the lips. Lite is small and fast, Pro is larger and looks better, and a separate upscaler brings clips to Full HD. It is free to use, including commercially, under the MIT license.
- Can I use Kandinsky 6.0 Video commercially?
- Yes. Kandinsky 6.0 Video is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Kandinsky 6.0 Video open source?
- Kandinsky 6.0 Video is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Kandinsky 6.0 Video on my own computer?
- Yes. The smallest variant, Kandinsky 6.0 Video Lite (50 steps), needs about 16 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Kandinsky 6.0 Video released?
- Kandinsky 6.0 Video was released by Kandinsky Lab on October 6, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Kandinsky 6.0 repository: README, updates and timings (GitHub) github.com
- Kandinsky-6.0-Pro-5s-Diffusers model page (Hugging Face) huggingface.co
- Kandinsky Video product page (Kandinsky Lab) kandinskylab.ai
- Kandinsky Lab news, 6 October 2026 release kandinskylab.ai
- Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation (arXiv 2610.05608) arxiv.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| LTX-2.5 LTX (Lightricks) | 22B | – | LTX-2.x Community (Conditional) | 16 GB Laptop or phone | |
| Wan 2.2 Alibaba Wan | 5B – 27B 14B active | – | Apache 2.0 (Permissive) | 24 GB Gaming GPU | |
| Cosmos 3 NVIDIA | 3.86B – 64.6B | – | OpenMDW-1.1 (Permissive) | 11 GB Laptop or phone | |
| MiniMax H3 MiniMax | 33.1B | – | MiniMax H3 Community (Conditional) | 166 GB Big-memory desktop |