Sber releases Kandinsky 6.0 Image Pro
Sber has released Kandinsky 6.0 Image Pro, a single model for generating and editing images that runs more than 40% faster than its predecessor and that Sber says performs on par with Flux 2 Max.
Sber's Kandinsky team announced Kandinsky 6.0 Image Pro on Habr. The model handles both text-to-image generation and image editing. It follows the Kandinsky 5 line, which included Text-to-Image and Image Editing Lite models and opened to the public at the AI Journey conference late last year.
The team says editing got the most attention in this release. In its own comparisons, the model performs at the level of Flux 2 Max and beats GPT Image 1.5. The post does not give the benchmarks, the evaluation method or any figures, and no one outside Sber has checked the results.
Sber says the model is more than 40% faster than the previous version. It credits three changes: a move to a mixture-of-experts (MoE) architecture, better parallelism during inference and optimizations to the attention mechanism.
The release also adds Image RAG. When a user sends a request, the system searches a knowledge base and adds relevant reference images to the context when they are needed. Sber says this improves how well the model renders Russian cultural subjects, such as Mezen painting and the traditional peaked cap known as a kartuz. It also lets the team add new knowledge without retraining the whole model. Image RAG works in both generation and editing modes.
Listed uses include removing objects and text, swapping objects, stylizing portraits, restoring and colorizing old photos, and designing building exteriors and interiors. Kandinsky already powers photo restoration in the "Immortal Regiment Online" service.
The model is available on GigaChat's Telegram bot, the Max messenger, giga.chat and the Android app. Sber says editing with up to three reference images and brush-based local editing are coming soon. The post does not mention open weights or API pricing.