Kandinsky 6.0 Video: Local AI Video with Audio on RTX 4090

👁 112 reads

Kandinsky 6.0 Video is an open model that generates a five-second clip with audio in a single run, including speech, sound effects and music. The Kandinsky Lab team released it on October 6, 2026, under the MIT license, which also allows commercial use. I ran both versions, Lite (3B) and Pro (29B), in ComfyUI on an RTX 4090. Lite takes about 2 minutes per clip; Pro takes about 6.5 minutes. Pro speaks English without errors. In my Polish speech test, it got 7 of 9 words right.

What is Kandinsky 6.0 Video?

Kandinsky is a family of models from Russia’s Sber, developed by Kandinsky Lab. Version 6.0 is the first open model in the family to generate video and audio at the same time. It has separate video and audio streams connected so that, for example, the sound of a wave crashing lines up with the frame in which the wave hits a rock. This also lets the model synchronize lip movements with speech.

According to the authors’ report, Pro outperforms the open LTX-2.5 on most VABench metrics. Gearbrain says it ranks behind only MiniMax H3 among open models. Those are the authors’ measurements and other people’s coverage, so I tested it on my own computer.

My RTX 4090 test

Hardware: RTX 4090 24 GB, 64 GB RAM, Windows 11, ComfyUI 0.41.0, models stored on an external USB SSD. I used the official template settings: 10 steps, CFG 1, PiFlow sampler, 864×480, 121 frames. I turned off the prompt expander, a separate Qwen 3.5 9B model (19 GB), and wrote full prompts directly. I did not enable Super Resolution. I measured from submitting the job to the saved MP4, including model loading. VRAM usage is the peak for the whole GPU, measured with nvidia-smi.

Pro int8 is 30.5 GB, so it does not fit on a 24 GB GPU. ComfyUI keeps some weights in RAM and transfers them as needed; it reported 29 GB as “staged.” That is why Pro is more than three times slower than Lite and needs 64 GB of RAM.

Speech: does it say what I wrote?

You enter the line to be spoken in the scene description between the <S> and <E> markers. I ran every clip through Whisper large-v3 and compared the transcription with the input. The Polish lines and transcriptions below are translated from the original Polish test material; the test itself was conducted in Polish. The English line was tested in English.

Although the authors’ materials do not mention Polish, Pro gets almost the entire sentence right, with a clear accent. It drops small words such as “z” (“from”) and mispronounces an ending, saying “miech” instead of “miód” (“honey”). That could work for a short line in an ad, but it is not ready for longer Polish dialogue.

Clips

All clips are unedited, with audio straight from the model. Turn your speakers on.

What I noticed

• Pro looks like a film frame. It adds cinematic black bars (2.39:1), soft lighting and natural-looking faces. Lite looks more like an ad and follows the prompt more literally.
• Lite follows the prompt more closely. For the lighthouse, it added a beam of light and splashing water. Pro produced a pretty but calmer shot without a beam or rain.
• Both versions match the sound to the visuals. Market noise, the hiss of a milk frother and crashing waves fit what you see. The background music requested in the barista prompt plays quietly, like it would in a café.
• Text in the frame is garbled, as it is with most video models. Lite put an unreadable sign on the espresso machine.

How to run Kandinsky 6 in ComfyUI

1. Update ComfyUI to version 0.38 or newer (I used 0.41).
2. In ComfyUI Manager, search for kandinsky6 (publisher: kandinskylab) and install it. Optionally, add kandinsky6-sr for upscaling. To install manually, copy the comfyui folder from the kandinskylab/kandinsky-6 repository to custom_nodes/kandinsky6, then install requirements.txt using ComfyUI’s Python.
3. Download a model. Lite distill (6.4 GB) goes in models/diffusion_models/Kandinsky-6.0-Lite-distill-5s-Diffusers/transformer/ alongside config.json. The official Pro distill in bf16 is 60.3 GB. I used the int8 version (30.5 GB) from kabachuha/Kandinsky-6.0-Pro-distill-5s-int8-convrot. Its author warns that it may not work with the authors’ node; it worked for me.
4. Download the shared components: the qwen_2.5_vl_7b.safetensors encoder and clip_l.safetensors (put them in text_encoders), hunyuan_video_vae_bf16.safetensors (in vae), the v1-44.pth audio VAE from MMAudio, and the 44 kHz BigVGAN vocoder (in audio_vae). The node README has the exact download links.
5. Open Workflow, then Browse Templates, then kandinsky6, and choose “Text to Video+Audio” or “Image to Video+Audio.”
6. If you turn off the prompt expander, write a full prompt in English: who is there, what they are doing, the camera and lighting; in the second field, describe the sound. Put spoken lines between <S>text<E>.

On Windows, launch ComfyUI with PYTHONUTF8=1 set. Without it, the nodes would not load on my setup because Polish characters appeared in the logs. See my general ComfyUI setup guide.

Kandinsky 6 compared with other open video models

For a broader comparison of open video models, including MAGI-2 and LTX-2.5, see Open video models in 2026. For more releases from October, including image and music models, see New local AI models.

The easiest option: Hangar

Both Kandinsky 6 versions are listed in Hangar, my free local model manager for Windows, with no account or subscription required. The catalog lists the models in Diffusers format: Lite distill is 26.7 GB in total, and Pro distill is 80.6 GB, including a 60.3 GB transformer. Choose Lite for a 24 GB GPU; Pro only fits with offloading to RAM. The MIT license allows commercial use, including in the EU. Kandinsky needs a separate engine in Hangar, which is still in development.

1. Download the 14 MB installer from Hangar. It is not digitally signed yet, so Windows SmartScreen will show a warning: click “More info,” then “Run anyway.”
2. Hangar detects your GPU (Nvidia with CUDA, from 8 GB VRAM; 12 GB or more is recommended) and shows whether the model will fit.
3. Find Kandinsky 6 Lite or Pro in the catalog to see the file size, license and whether the model can be used in the EU.
4. For now, generate in ComfyUI with the kandinsky6 node, following the instructions above. Hangar can install ComfyUI portable from scratch (it took 231 seconds on my setup) or connect to an existing ComfyUI installation.

Hangar lists the model, but generation through Hangar is not working yet (as of October 17, 2026). For now, use ComfyUI as described above. Read more about the tool at Hangar.

Frequently asked questions

How much VRAM does Kandinsky 6 need?

At 864×480, Lite used a peak of 18.8-19.9 GB across the whole GPU. Pro int8 used 21.4 GB, but only because the remaining weights (about 29 GB) were kept in RAM. Lite may run on a 16 GB GPU with offloading to RAM, but more slowly. I do not recommend Pro without 64 GB RAM and 24 GB VRAM.

Can Kandinsky 6 speak Polish?

Pro does reasonably well: it got 7 of 9 words right in my test sentence, including a clear translation of the Polish original: “Good morning” and translation of the Polish original: “Apples straight orchard”. Lite mangles most of the sentence. For spoken Polish lines, choose Pro. This was a Polish-language test, not an English retest of the translated example.

Can I use the clips commercially?

The MIT license for the weights and code allows commercial use and does not exclude any countries. Keep in mind that the int8 version I used comes from an independent author and is under the Apache 2.0 license, which also allows commercial use.

How can I make a clip longer than 5 seconds?

The 6.0 models are trained on five-second clips (121 frames). The node lets you enter a higher frame count, but I have not tested longer clips and cannot say how well the quality holds up. You can join longer scenes from several clips, for example by using the last frame as the starting image in image-to-video mode.

Lite or Pro?

Choose Lite for exploring ideas and shots without speech; it is more than three times faster. Choose Pro for a finished scene with dialogue or shots that should look cinematic.

Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →

Sources

• Kandinsky 6.0 Lite distill on Hugging Face (model card, MIT license)
• kandinskylab/kandinsky-6 on GitHub (code and ComfyUI node)
• Kandinsky 6.0 Video technical report (arXiv)
• Kandinsky 6.0 Pro distill int8 (kabachuha)
• My RTX 4090 measurements, October 17, 2026.