Kandinsky

Kandinsky is the open-weight counterexample in this category: a family of text-to-image models published by the ai-forever organisation, whose authorship sits with Sber-affiliated researchers, released under Apache 2.0 and documented as a three-part stack — a Flan-UL2 text encoder, a latent-diffusion U-Net and a MoVQ encoder/decoder. The current line in the publisher's own repository is Kandinsky 3.0 with a 3.1 Flash variant that swaps half of the U-Net encoder for additional head predictions to run faster. Because the weights are downloadable and the licence is permissive, the model is yours to keep: fine-tune it, run it offline, ship it inside a product, with no per-image meter between you and the GPU.

Two caveats should decide a purchase. Language: the repository's own prompt beautifier instructs the model to expand inputs in English, and the training data is documented as carrying deliberate references to Russian culture — so the multilingual reputation is about the corpus more than about writing good prompts in whatever language you prefer. Availability: the hosted front end this listing points at, FusionBrain, could not be reached while this review was written, so only the model release itself could be verified today. If you want a polished browser generator with free daily credits, use a hosted service and keep the weights for the work where licence freedom, offline execution or a research baseline is the requirement.