whatmodel.sh / video / Kandinsky 6.0 Video Pro
Kandinsky 6.0 Video Pro
↗ share
+ contribute
Kandinsky 6.0 Video Pro is an open-weights video model from Sber, released October 2026. Its published size is 29B parameters. Licence: mit.
Catalogue identity notes
Kandinsky 6.0 Video (Sber / Kandinsky Lab), open-sourced 2026-10-06 under MIT (GitHub kandinskylab/kandinsky-6 README news, @8195e72). The arXiv report (2610.05608v1) evaluates 'Kandinsky 6.0 Video Pro and Kandinsky 6.0 Video Lite after RL post-training'; the family table on the HF cards lists separate Pro / Pro-distill (10 steps) / Pro-pretrain and Lite / Lite-distill / Lite-pretrain repos, and the non-distilled, non-pretrain repo is described as 'the main ... checkpoint, the strongest quality setting' of its line, so it is mapped to the report's final (RL post-trained) model. Distilled and pretrain checkpoints are not catalogued (no scored rows). Pro repo kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers pinned 295ca97 (gated: Hugging Face asks users to agree to share contact information; licence MIT in card metadata). Text-to-audio-video and image-to-audio-video, 5-second clips with 44 kHz audio.
Model information sources
Release date ↗ · Weights licence ↗ · Parameter count ↗
29B in total as stated in the report (video stream 19B, audio stream 5B, cross-attention 5B); Qwen2.5-VL/CLIP text encoders, HunyuanVideo VAE and the MMAudio VAE/vocoder are separate components.
$ how it compares
The video models ranked either side of Kandinsky 6.0 Video Pro, out of 139 with a score, then more from Sber.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 75 |  Cosmos-Predict2-14B-Video2World | 48±8 | comparison only | ||
| 76 |  MAGI-1 24B | 48±4 | comparison only | ||
| 77 | Kandinsky 6.0 Video Pro | 48±7 | local | ||
| 78 |  Wan2.1-I2V-14B-720P | 48±4 | comparison only | ||
| 79 | Cosmos-Predict2-2B-Video2World | 48±8 | comparison only | ||
| more from Sber | |||||
| 57 | Kandinsky 5.0 Pro | 52±2 | api $0.040/sec +local | ||
| 64 | Kandinsky 6.0 Video Lite | 51±7 | local | ||
| 82 | Kandinsky 5.0 T2V Lite | 47±2 | local | ||
$ what it's good at
text to video█····█·····█ #3 of 3 * · 77% of its score · 1 test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| VABench T2AV micro Visual QA lab | 65.2% ✎ | ▄█··▄·▄···▄█ beats 2/7 | 77% |
$ where to run it
Other ways to run it
- Run it as published · needs ~108.5 GB, NVIDIA only · from huggingface.co/kandinskylab
uv run --with git+https://github.com/huggingface/diffusers --with torch --with transformers --with accelerate --with av python -c "import torch; from diffusers import Kandinsky6TI2VAPipeline; from diffusers.utils import encode_video; p=Kandinsky6TI2VAPipeline.from_pretrained('kandinskylab/Kandinsky-6.0-Pro-5s-Diffusers', torch_dtype=torch.bfloat16); p.enable_model_cpu_offload(); o=p(prompt='A red dragon wakes on a mound of gold coins in a sunlit cave and slowly unfolds its wings. Audio: deep rumbling breath, clinking coins.', height=480, width=864, num_frames=121, num_inference_steps=50, guidance_scale=5.0); encode_video(o.frames[0], fps=24, output_path='output.mp4', audio=o.audio[0][None], audio_sample_rate=p.audio_sample_rate)"
Weights and documentation: huggingface.co/kandinskylab ↗
$ how the score was worked out · Kandinsky 6.0 Video: Foundati… 77%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 Kandinsky 6.0 Video: Foundation Models for Synchronized Video and Audio Generation (arXiv:2610.05608v1) · lab-reported | 1 | 77% |
| prior · pull toward the middle | 23% | |
| weighted average | = 48.2 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json