whatmodel.sh / video / Kandinsky 5.0 Pro
Kandinsky 5.0 Pro
↗ share
+ contribute
Kandinsky 5.0 Pro is an open-weights video model from Sber, released November 2025. Its published size is 19B parameters. Licence: mit.
Model information sources
Release date ↗ · Weights licence ↗ · Parameter count ↗
19B as stated for the Video Pro line-up (T2V and I2V variants); README describes the DiT as the main generative module with Qwen2.5-VL + CLIP text encoders and HunyuanVideo VAE, and does not say whether 19B includes the encoders, so read it as the Pro generative model size.
$ how it compares
The video models ranked either side of Kandinsky 5.0 Pro, out of 139 with a score, then more from Sber.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 55 |  HunyuanVideo 1.5 (I2V) | 53±2 | api $0.075/sec +local | ||
| 56 |  Kling 2.1 Master | 53±2 | api $0.28/sec | ||
| 57 | Kandinsky 5.0 Pro | 52±2 | api $0.040/sec +local | ||
| 58 |  HY-OmniWeaving | 52±4 | comparison only | ||
| 59 | HunyuanVideo 1.5 | 52±2 | api $0.075/sec +local | ||
| more from Sber | |||||
| 64 | Kandinsky 6.0 Video Lite | 51±7 | local | ||
| 77 | Kandinsky 6.0 Video Pro | 48±7 | local | ||
| 82 | Kandinsky 5.0 T2V Lite | 47±2 | local | ||
$ what it's good at
text to video▆▄▃▄▆▅█▃▂▂▂▄ #28 of 68 * · 98% of its score · 2 tests
Beats Kling 2.1 Master here, though they rank higher overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| LMArena video | 1175.2 ✎ | ▂▂▃▆█▆·▆▅▁▂▇ beats 11/41 | 95% |
| HumanScore (T2V) | 86.5% ✎ | ▂··▂··█▄▂··▆ beats 7/11 | 3.1% |
$ where to run it
Through an API
api · fal.ai · $0.040/sec · fal.ai ↗ · price sourceUSD per video second at 512P (endpoint default) on fal.ai. Product: fal Kandinsky5 Pro text-to-video endpoint. API model: fal-ai/kandinsky5-pro/text-to-video (5-s clips). Price basis: rate printed by fal; 1024P $0.12 per video second. No Sber USD API tariff was used. Currency: USD. fal.ai model page, observed 2026-09-25.
Other ways to run it
- Run it as published · needs ~75.7 GB, NVIDIA only · from huggingface.co/kandinskylab
uv run --with torch --with diffusers --with transformers --with accelerate --with imageio-ffmpeg python -c "import torch; from diffusers import Kandinsky5T2VPipeline; from diffusers.utils import export_to_video; p=Kandinsky5T2VPipeline.from_pretrained('kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s-Diffusers', torch_dtype=torch.bfloat16); p.transformer.set_attention_backend('flex'); p.enable_model_cpu_offload(); export_to_video(p(prompt='A cat and a dog baking a cake together in a kitchen.', negative_prompt='Static, 2D cartoon, cartoon, 2d animation, paintings, images, worst quality, low quality, ugly, deformed, walking backwards', height=768, width=1024, num_frames=121, num_inference_steps=50, guidance_scale=5.0).frames[0], 'output.mp4', fps=24, quality=9)"
Weights and documentation: github.com/kandinskylab ↗
$ how the score was worked out · LMArena leaderboard dataset 95%, HumanScore: Benchmarking Huma… 3%
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 |
|---|---|---|
 LMArena leaderboard dataset | 1 | 95% |
 HumanScore: Benchmarking Human Motions in Generated Videos | 1 | 3% |
| prior · pull toward the middle | 2% | |
| weighted average | = 52.4 |
$ 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