Sber has rolled out Kandinsky 6.0 Image, the latest update to its flagship AI image generation and editing model embedded in the GigaChat assistant. Released April 28, 2026, this version runs up to twice as fast as before, understands complex prompts more precisely, and offers unlimited free usage. It’s a notable upgrade in Russia’s growing AI scene, competing with international models like Nano-Banana and Qwen in certain tasks.
- Kandinsky 6.0 Image is built on a Mixture of Experts (MoE) architecture and features real-time image retrieval with Image RAG technology.
- According to Sber’s internal benchmarks, the model’s performance rivals other recent AI image generators such as Nano-Banana, Flux, and Qwen.
- GigaChat users have generated hundreds of millions of images through the platform since launch.
New AI editing features in Kandinsky 6.0
The update introduces several advanced editing features. Photo restoration sharpens faded or damaged images, while stylization transforms photos into anime, comic book, or cartoon-like illustrations. The new ”neuro photoshoot” mode can generate multiple frames from a single image, swapping backgrounds, outfits, or locations.
Additional tools include retouching, virtual makeup application, style changes, and filter overlays. Users can now edit details more precisely – for example, changing a character’s hairstyle or adding glasses without affecting other parts of the image. The model also shows improved detail in complex scenes such as interiors, table settings, and group photos, with fewer errors when generating text on images.
How Image RAG improves real-time visual reference retrieval
Kandinsky 6.0’s built-in Image RAG system lets the AI search for visual references in real time. This means it can generate images featuring new pop culture characters or objects missing from its original training data by dynamically finding matching references online. This approach overcomes a common limitation of AI models trained on fixed datasets, allowing Kandinsky to stay visually current.
Mixture of Experts architecture speeds up Kandinsky 6.0
Kandinsky 6.0 is powered by Mixture of Experts architecture, which activates only relevant expert modules for each request instead of running the entire neural network. This selective computation speeds up response times, lowers resource use, and makes targeted improvements easier for developers.
Anton Frolov, Senior Vice President of Sber and head of Generative AI Development, said: ”Image editing and creation are among the most popular features in our AI assistant GigaChat; users have already generated hundreds of millions of pictures. Our goal is to provide everyone with a professional-grade tool offering versatile capabilities-from creative projects to professional work. Kandinsky 6.0 marks a significant step toward that.”
Frolov also revealed that video capabilities are on the horizon for GigaChat users. Kandinsky Video integration already lets users animate generated images, indicating future expansions into video editing and generation.
Kandinsky 6.0 Image is available now via the GigaChat web interface, mobile app, and messaging platforms.
Russia’s Sberbank leads one of the region’s most ambitious AI projects. With Kandinsky, Sber aims to compete with global offerings from Apple, Google, and Meta by delivering accessible tools that combine image generation with advanced editing. The use of a retrieval-augmented approach (Image RAG) is especially notable, as it helps keep AI art generation more current and adaptable than most Western competitors, which often rely strictly on static training sets.
Looking ahead, watch how quickly Sber integrates video editing tools and whether Kandinsky’s MoE architecture will influence similar projects worldwide. As AI art and editing tools become mainstream, hybrid models combining generation with real-time reference search could set a new standard for visual AI assistants.