Kandinsky

Kandinsky today is really two products wearing one name. The first is Kandinsky 6.0 Image Pro, released in April 2026 and described by Sber as the current flagship: a single model that handles both text-to-image generation and instruction-based editing — object removal, restyling, photo restoration and colorization, interior and exterior design mockups — built on a Mixture-of-Experts architecture that Sber says runs more than 40% faster than its 5.0 predecessor. It also includes a retrieval-augmented feature Sber calls Image RAG, which pulls reference images from a knowledge base to improve accuracy on Russian cultural motifs, characters, and art styles. None of this is available for local, offline use: as of August 2026, Kandinsky 6.0 Image Pro has not appeared as downloadable weights anywhere, including Hugging Face. It's reachable only through GigaChat, Sber's ChatGPT-style assistant, via a Telegram bot, the Max messenger, the giga.chat web app, and an Android app.

The second product is the open-weight side of the family: Kandinsky 5.0, which shipped through late 2025 as a 6-billion-parameter Image Lite/Editing pair plus 2B and 19B video models (Video Lite and Video Pro). These are real, freely downloadable checkpoints — supervised fine-tuned and base "pretrain" variants for both image and video — with example notebooks, ComfyUI integration, and official support in Hugging Face's Diffusers library. Sber's own tracking claims Kandinsky 5.0 Video Pro ranked first among open-source models on LMArena's Text-to-Video Arena leaderboard in December 2025, a vendor claim worth treating with the usual skepticism but one that at least points to a model taken seriously by outside benchmarkers.

This is the detail that separates Kandinsky from most closed image APIs, and it's worth being precise about which versions actually qualify. Kandinsky 2.0 through 3.1 are published on Hugging Face under the kandinsky-community and ai-forever organizations, licensed Apache-2.0 — genuinely open, genuinely old (the newest of that batch dates to April 2024). Kandinsky 5.0's Image, Video Lite, and Video Pro checkpoints are published under the kandinskylab organization on Hugging Face and licensed MIT, one of the most permissive open-source licenses available, with no restriction on commercial use beyond attribution. There's visible evidence of real international pickup: independent Hugging Face users and studios — including Runware, a European inference provider, and community contributors packaging GGUF-quantized and ComfyUI-ready versions — have built on Kandinsky 5.0 without any apparent friction from Kandinsky's Russian origin, because downloading a model from Hugging Face doesn't route through Sber's consumer infrastructure at all.

What isn't open is the part most readers actually want: the current flagship. Kandinsky 6.0 Image Pro sits behind GigaChat's hosted apps only. That's a meaningful shift from the project's earlier "open by default" posture, and it means the honest answer to "are Kandinsky's weights open?" is "the previous generation, yes — the model Sber is actually promoting right now, no."

Access to the flagship runs entirely through GigaChat's consumer surfaces, and Sber describes generation and editing there as free, with no published usage caps in English-language materials. A paid GigaChat Pro subscription has historically existed alongside the free tier, priced in rubles, offering higher limits — but current pricing isn't published anywhere in English as of this writing, so treat any specific figure as unverified. Developers can also reach GigaChat and related Sber Cloud services through a usage-based API, though sign-up and documentation are built for a Russian-speaking developer audience.

The other path costs nothing but compute: download a Kandinsky 5.0 checkpoint from Hugging Face and run it on your own GPU, or rent cloud compute from any provider. That route sidesteps GigaChat entirely, at the cost of using a model generation behind the current flagship.

Stable Diffusion is the fairest openness comparison — both projects publish real weights researchers and hobbyists can run locally, and both maintain active open-source ecosystems (LoRAs, ControlNet variants, community fine-tunes). Kandinsky's differentiator inside that comparison is its training emphasis: 10 million-plus images of Russian text in different styles feed its Cyrillic rendering, an area where Stable Diffusion and most Western open models are noticeably weaker.

Midjourney and DALL-E sit on the other side of the comparison — closed, hosted, and optimized for English-language prompting and broad aesthetic polish. Sber's own side-by-side testing claims Kandinsky 6.0 Image Pro performs on par with Flux 2 Max and ahead of GPT Image 1.5 specifically on editing tasks; that's a vendor claim from Sber's own blog post announcing the model, not an independently reproduced benchmark, and it should be read that way. For an English-prompting user with no specific reason to touch the Kandinsky ecosystem, Midjourney or DALL-E remain the more practical picks — faster to sign up for, better documented in English, and not tied to a single national payments and messaging ecosystem.

The practical accessibility problem is the biggest one for an international audience, and it's separate from output quality. GigaChat's apps are Russian-market products: the sign-up flow, terms of service, and support documentation are primarily in Russian, and Sberbank — Kandinsky's ultimate parent — has been under US and EU sanctions since 2022, which complicates payment processing and account setup for many people outside Russia regardless of the model's technical merits. None of that affects the open-weight 5.0 checkpoints, which anyone can pull from Hugging Face, but it does mean the model Sber is actively promoting as its best work is effectively closed off to casual international users.

Anyone who needs strong English-prompt output, straightforward billing, or clear Western commercial-licensing terms should look elsewhere. Anyone specifically interested in Cyrillic image generation, Russian-cultural visual references, or an MIT-licensed diffusion model to self-host should take Kandinsky 5.0 seriously — it's a legitimate, actively developed alternative to Stable Diffusion, not a curiosity.

Kandinsky is a genuinely capable, well-published research program that happens to be split across two very different access models: an open, MIT-licensed previous generation that anyone can download and run, and a closed, Russia-first hosted flagship that most international readers won't be able to use comfortably even if they wanted to. Judge it by which half applies to you.