Image models · Sber AI

Kandinsky 3

Name: Kandinsky 3
Author: Sber AI

Backbone: 3.0B
Native output: 1024×1024
Default steps: 50
License: apache-2.0
Released: 2023-11

How to run it

Original weights · Hugging Face

Full-precision files from Sber AI, for fine-tuning, vLLM, Transformers, Diffusers or ComfyUI.

huggingface.co/kandinsky-community/kandinsky-3 (opens in a new tab)

No sign-in needed.

Official repo · GitHub

Reference code, inference scripts and docs from Sber AI. The weights themselves download from Hugging Face.

github.com/ai-forever/Kandinsky-3 (opens in a new tab)

git clone https://github.com/ai-forever/Kandinsky-3

Links checked 7 days ago against Hugging Face, Ollama and GitHub.

Apps that run Kandinsky 3

Our pick for each kind of computer, assuming the model fits in its memory. Pick your own GPU or device in the Model Finder for an exact answer.

Mac (Apple silicon)

ComfyUI

Runs this model with built-in or popular custom nodes; GGUF files load with ComfyUI-GGUF.

Also good:

Windows PC, NVIDIA GPU

ComfyUI

Runs this model with built-in or popular custom nodes; GGUF files load with ComfyUI-GGUF.

Also good:

Linux, NVIDIA or AMD GPU

ComfyUI

Runs this model with built-in or popular custom nodes; GGUF files load with ComfyUI-GGUF.

Also good:

App	Type	Opens it as	Systems	Engine	Start
ComfyUI	Desktop app	Built-in support	macOSWindowsLinux	PyTorch, GGUF via ComfyUI-GGUF	Get ComfyUI
SwarmUI	Desktop app	Built-in support	macOSWindowsLinux	ComfyUI	Get SwarmUI
SD.Next	Desktop app	Built-in support	macOSWindowsLinux	PyTorch	Get SD.Next
Stability Matrix	Desktop app	Built-in support	macOSWindowsLinux	installs ComfyUI, Forge, InvokeAI, SwarmUI, SD.Next or Fooocus	Get Stability Matrix

Which apps can open a model depends on the files published for it: GGUF for llama.cpp-based apps, MLX for Apple’s engine, the original weights for vLLM and SGLang. Compare every app.

Memory needed

Peak GPU memory at the native output size. Offloading moves the text encoder, or the whole pipeline, to system RAM between steps.

Precision	Everything on GPU	Text encoder offloaded	Sequential offload	Download
`FP16`	25.9 GB	8.7 GB	4.2 GB	Original weights (opens in a new tab)

Components

Backbone: 3B
VAE: 0.27B
Flan-UL2 encoder: 8.6B

Best hardware for Kandinsky 3

The least expensive GPU in our catalog that runs it without offloading is the Intel Arc A380 (launch price $139) at FP16.

Quality	Hardware	Quant	Memory	Speed	Runs
48Fair quality (C) · 48 of 100: how capable the model is at FP16, whatever it runs on. Match score 76/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 5090	`FP16`	25.9 GB / 32.0 GB	7.5 s	Runs great
48Fair quality (C) · 48 of 100: how capable the model is at FP16, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 4070 SUPER	`FP16`	8.7 GB / 12.0 GB	18 s	Runs wellencoder in RAM
48Fair quality (C) · 48 of 100: how capable the model is at FP16, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA RTX PRO 6000 Blackwell	`FP16`	25.9 GB / 96.0 GB	6.6 s	Runs great
48Fair quality (C) · 48 of 100: how capable the model is at FP16, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking.	AMD Radeon RX 9070 XT	`FP16`	8.7 GB / 16.0 GB	13 s	Runs greatencoder in RAM
48Fair quality (C) · 48 of 100: how capable the model is at FP16, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 5080	`FP16`	8.7 GB / 16.0 GB	13 s	Runs greatencoder in RAM
48Fair quality (C) · 48 of 100: how capable the model is at FP16, whatever it runs on. Match score 67/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 4080 SUPER	`FP16`	8.7 GB / 16.0 GB	13 s	Runs greatencoder in RAM
48Fair quality (C) · 48 of 100: how capable the model is at FP16, whatever it runs on. Match score 66/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 4070 Ti SUPER	`FP16`	8.7 GB / 16.0 GB	15 s	Runs wellencoder in RAM
48Fair quality (C) · 48 of 100: how capable the model is at FP16, whatever it runs on. Match score 66/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 5070 Ti	`FP16`	8.7 GB / 16.0 GB	15 s	Runs wellencoder in RAM
48Fair quality (C) · 48 of 100: how capable the model is at FP16, whatever it runs on. Match score 64/100 (quality, fit, speed and memory use) sets the ranking.	NVIDIA GeForce RTX 4090	`FP16`	8.7 GB / 24.0 GB	9.5 s	Runs greatencoder in RAM
48Fair quality (C) · 48 of 100: how capable the model is at FP16, whatever it runs on. Match score 64/100 (quality, fit, speed and memory use) sets the ranking.	AMD Radeon RX 7900 XTX	`FP16`	8.7 GB / 24.0 GB	11 s	Runs greatencoder in RAM

Check it on your own hardware

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.