Kandinsky LabKandinsky Image
Kandinsky 4.1 Image
Sber announced this image model on 5 June 2025. It replaced the earlier U-Net approach with a diffusion transformer and applied supervised fine-tuning to carefully selected artistic data.
Access not confirmed
Expert selection shaped the fine-tuning data
The training process used specialist image review and separately trained domain LoRAs, then combined their weights. Guidance distillation reduced inference work.Sber: Kandinsky 4.1 Image, 5 June 2025
The announcement documents historical product access
The launch named GigaChat and Kandinsky bots. Those product names do not prove that 4.1 remains separately selectable today; downloadable weights were not verified.Sber: Kandinsky 4.1 Image, 5 June 2025
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Access not confirmed |
| Launch date |