Kandinsky 6.0 Pro, distilled to 10 PiFlow steps: text (and optionally a first-frame image) to a 5 s, 24 fps video with audio. A Qwen3.5-9B beautifier expands the prompt first (can be switched off under Advanced settings): staranov's plain beautifier (beautifier_os_plain/, system prompts for t2av and i2av), which writes the description in the language of the request.

• Weights: kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers, pinned to the revision in app.py. Needs the HF_TOKEN secret with read access to it.
• Code: the diffusers/ package next to app.py (Kandinsky 6 branch, with bf16 LayerNorm/addcmul speedups in transformer_kandinsky6.py) is used instead of the PyPI one.
• Hardware: ZeroGPU large (half an RTX PRO 6000). The DiT streams block by block from RAM; SageAttention on the video self-attention. About 275 s per request.