Kandinsky 3.0
Parameters
2.1B
Modality
Image
Auxiliary Parameters
4.7B
Architecture
Diffusion Transformer
License
Apache 2.0
Release Date
5 Dec 2023
API Pricing (per 1M)
Self-hosted only
Hardware Sizing
Precision
Workload Memory: 3.67 GB
INT4 · 1024x1024
DiT Backbone Weights: 0.98 GB
VAE Decode Workspace: 0.19 GB
Runtime & Graphs: 2.50 GB
Minimum Required Hardware
1x NVIDIA L4 24GB
3.7 / 21.6 GB usable
1x RTX 3060 12GB
3.7 / 10.8 GB usable
Apple M4 24GB
3.7 / 18.0 GB usable
Determine requirements for your specific workload:
Open in VRAM Calculator
Architecture Diagram
About Kandinsky 3.0
Kandinsky 3.0 is a 2.1B parameter latent diffusion model powered by a pure UNet backbone and Flan-UL2 text encoder, optimized for photorealism and multilingual prompt processing.
Technical Specifications
Attention
Attention Heads
64
Attention Head Dimension
64
Dimensions
Auxiliary Parameters
4.7B
Hidden Dimension Size
4,096
Number of Layers
12