Kandinsky AI: what Sber's Kandinsky 6.0 and open models can do

What is Kandinsky AI

Kandinsky is Sber’s family of image and video generation models. It’s built by the Sber AI team, branded Kandinsky Lab since 2025, together with the AIRI institute. The first version came out in 2022, and Sber has shipped a new generation roughly once a year since then.

Most articles still describe Kandinsky 2.2 or 3.1 and the Fusion Brain website. The 2026 lineup looks different:

Version	Date	What’s new
Kandinsky 2.1	Apr 2023	The model that made Kandinsky widely known
Kandinsky 3.0 and 3.1	Nov 2023, Apr 2024	Faster generation, inpainting, 4K upscaling
Kandinsky 4.0	Dec 2024	Video up to 12 seconds in HD with an audio track
Kandinsky 4.1 Image	Jun 2025	DiT architecture, 4× faster generation
Image update	Oct 2025	Cyrillic text without add-on modules
Kandinsky 5.0	Nov 2025	Open image and video models under MIT
Kandinsky 6.0 Image	Apr 28, 2026	Flagship for images and photo editing in GigaChat
Kandinsky 6.0 Video	Oct 6, 2026	Video with sound and lip sync

What Kandinsky can do

Photo editing. Sber pitches Kandinsky 6.0 Image as a free editing tool: restoring old photos, restyling, AI photo shoots, changing outfits and locations, retouching. Targeted edits, such as a new hairstyle or added glasses, leave the rest of the image untouched, according to the developers.

Merging up to three photos. Since June 9, 2026, GigaChat lets you upload three photos and combine them into one composition: the model picks out a character, object, background or style. The same update added editing of a selected area and automatic aspect ratios: vertical for portraits, horizontal for panoramas, square for avatars.

Reference search. If the model doesn’t know what something looks like, it searches the web for images of it. That helps with things that appeared after the model was trained.

Cyrillic text. Since October 2025, Kandinsky renders Russian text without separate modules: the model was fine-tuned on 10 million images with Russian lettering. The developers admit long text and mixed Cyrillic and Latin may take a few attempts.

Video with sound. Kandinsky 6.0 Video generates a 5-second clip together with its audio: lip-synced speech, music and ambient noise. The model renders at 864×480, and a separate upscaler takes it to Full HD. Details are in our news post on Kandinsky 6.0 Video.

Russian cultural context. Sber stresses that Kandinsky understands Russian concepts well. Global models handle this too: the samovar, glass holders and tiled stove in the image below were drawn by Nano Banana Pro from an English prompt.

Open Kandinsky 5.0 models

In November 2025, Sber released the Kandinsky 5.0 family on GitHub and Hugging Face under the MIT license, which allows commercial use:

• Image Lite — 6B parameters, 1K images (1280×768, 1024×1024), strong text rendering;
• Image Editing — the same model for photo-based edits;
• Video Lite — 2B parameters, 5- and 10-second clips, runs on a 12 GB GPU;
• Video Pro — 19B parameters, up to 10 seconds in HD (1280×768).

In December 2025, Video Pro was the top open text-to-video model on the Arena leaderboard. The models are supported in Diffusers. Kandinsky 6.0 Video weights are open too, while 6.0 Image runs only in GigaChat.

Where to use Kandinsky for free

• GigaChat — the giga.chat website: images, photo edits and video in a single chat.
• GigaChat bots in Telegram and MAX — the same features on your phone.
• Your own computer — the open 5.0 and 6.0 Video models; you’ll need a GPU.

GigaChat requires a Sber ID, which you can get without a Sberbank card. The service only works from a Russian IP address, so it won’t open from abroad or through a VPN.

What Kandinsky is good for

• Greeting cards, social media illustrations and Russian-themed scenes — free and unlimited.
• Quick edits of your own photos — restoration, background and outfit swaps, retouching.
• Short clips with sound for social media — Kandinsky 6.0 Video.
• Building your own service or fine-tuning — the open 5.0 models under MIT.

Kandinsky vs. alternatives

	Kandinsky 6.0 Image	Nano Banana 2.1	GPT Image 2.5	Seedream 5.0 Pro
Arena rank, text-to-image (Oct 24, 2026)	not ranked	6	1 (Sunburst)	12
Reference images	up to 3	up to 14	several	up to 10
Cyrillic text in images	yes	yes	yes	yes
Price	free in GigaChat	paid	paid	paid
Access	Sber ID, Russian IP	Google API	OpenAI API	ByteDance API
Open weights	no (5.0: yes)	no	no	no

Video tells the same story: the open Kandinsky 5.0 Video Pro ranks 37th of 48 on the Arena leaderboard, far behind Veo 3.1 and Seedance. Kandinsky wins on price and open weights; global models win on quality, resolution and the number of references.

How to write prompts for Kandinsky

• Russian prompts work best. Kandinsky was trained on Russian data, so it doesn’t need a translation step; English works too.
• Name the style directly: “oil painting,” “anime,” “pixel art,” “studio photo.” GigaChat has these styles built in.
• Describe the whole scene: who, where, lighting, camera angle. Version 6.0 handles complex multi-part prompts.
• For edits, say what changes and what stays: “change the background to a winter forest; keep the face and clothes.”
• Keep text short and in quotes. The developers suggest regenerating long text or mixed Cyrillic and Latin several times.

How to use Kandinsky in AI2Face

Kandinsky is available in the AI2Face generator, with no Sber ID and no Russian IP required. You can run one prompt on Kandinsky, Nano Banana, GPT Image and Seedream and compare the results. The cost of each generation is shown before you run it, you can pay by card or crypto, and credits from packs don’t expire. New accounts get welcome credits after signing up. For all image models, see our guide to AI image generators.

Limitations

• Not on independent image leaderboards. There’s no Arena data to compare Kandinsky 6.0 Image with global models.
• Short videos. Kandinsky 6.0 Video is 5 seconds with no extension, and Full HD is upscaled from 480p.
• Sber’s own access is Russia-only. GigaChat requires a Sber ID and a Russian IP address.
• The flagship is closed. You can download 5.0 and 6.0 Video, but not 6.0 Image.
• Long text is unstable. The developers themselves recommend several attempts for long inscriptions.

Bottom line

Kandinsky AI is the strongest free option from a Russian company: Kandinsky 6.0 Image in GigaChat generates and edits photos with no cap, understands Russian and renders Cyrillic. That’s enough for greeting cards, social posts and quick edits. If an image is for an ad, a marketplace listing or a client, go with the top of the leaderboard — GPT Image 2.5 or Nano Banana 2.1 — and compare them with Kandinsky on the same prompt in AI2Face. The open Kandinsky 5.0 models under MIT are a solid base for your own service.