- Services
Kandinsky is Sber AI’s open-weight text-to-image and text-to-video diffusion family. Kandinsky 3 and Kandinsky Video are released under permissive licenses for self-hosting.
See all models from Sber AIModels in family
2
Open weight
2
API only
0
Avg score
38.0
Top benchmark
53.0
Video Arena
Total HF downloads
45
Primary modality
Image
First release
Sep 2025
Latest release
Nov 2025
Every release in the Kandinsky family, ranked by composite score across benchmarks, popularity, efficiency, and versatility.
| # | Model | Modality | Score | Params | Released |
|---|---|---|---|---|---|
| 1 | video | CC41.9 | 2B | Sep 2025 | |
| 2 | video | DD34.0 | 19B | Nov 2025 |
When each release shipped, newest first. Useful for tracking version cadence.
Nov 20
Sep 29
Composite grades across this family. Higher is better, blending benchmarks, popularity, and efficiency.
Models with downloadable weights, ranked by composite score.
| # | Model | Modality | Score | Params | Released |
|---|---|---|---|---|---|
| 1 | video | CC41.9 | 2B | Sep 2025 | |
| 2 | video | DD34.0 | 19B | Nov 2025 |
The Kandinsky family is a series of AI models from Sber AI. This page lists every release in the family with its benchmark scores, parameter count, and hardware requirements.
By composite score, Kandinsky 5.0 Video Lite is currently the top model in the family. For local inference, match the parameter count to your VRAM budget. For quality, pick the highest scorer that fits.
See the open-weight section above for models you can run locally. The API-only section lists closed releases that must be accessed through the provider’s API.
Spin up an instance in the cloud, or pick local hardware that fits.
Advertising disclosure: we earn commissions when you shop through the links below.
Vast.ai
Decentralized GPU marketplace with the lowest hourly prices.
RunPod
Pay-per-second GPU rentals starting at $0.20/hr.
Digital Ocean
Spin up a GPU droplet in minutes, starting at $0.75/hr.
Vultr
GPU cloud with hourly and monthly plans, starting at $0.50/hr.
GPU Mart
Dedicated GPU servers and VPS billed monthly, starting at $0.50/hr.
PPQ.ai
Multi-model inference gateway for production workloads.
Find Local Hardware
See which GPU, Mac, or workstation can run Kandinsky on-prem.
Local LLM Mini PCs
Compact machines that run Kandinsky on your desk. Chosen for local inference.
