K5-GGUF

GGUF bundles for Kandinsky 5 Pro 5s text-to-video, prepared for kandinsky.cpp (https://github.com/AkaneTendo25/kandinsky.cpp).

Contents

• q4/: self-contained q4_0 bundle
• q8/: self-contained q8_0 bundle

Each bundle contains:

• kandinsky5pro_t2v_sft_5s.*.gguf
• qwen.gguf
• clip.gguf
• vae3d.gguf
• vae.gguf

qwen.gguf and clip.gguf include the tokenizer data in GGUF metadata, so no external tokenizer files are required for the text path.

Official Sources

Converted from the official upstream model releases used by Kandinsky 5:

• DiT checkpoint: https://huggingface.co/kandinskylab/Kandinsky-5.0-T2V-Pro-sft-5s
• Qwen text encoder: https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct
• CLIP text encoder: https://huggingface.co/openai/clip-vit-large-patch14
• VAE / VAE3D source: https://huggingface.co/hunyuanvideo-community/HunyuanVideo

Upstream download script:

• https://raw.githubusercontent.com/kandinskylab/kandinsky-5/main/download_models.py

Conversion

These GGUF bundles were produced with the conversion script from the main kandinsky.cpp repository:

• code: https://github.com/AkaneTendo25/kandinsky.cpp
• script: https://github.com/AkaneTendo25/kandinsky.cpp/blob/main/scripts/convert_to_gguf.py

python scripts\convert_to_gguf.py `
  --model-dir <path-to-Kandinsky-5> `
  --output-dir .\models\q4 `
  --type q4_0 `
  --dit <dit_path> `
  --qwen <qwen_path> `
  --clip <clip_path> `
  --vae3d <vae3d_path>

python scripts\convert_to_gguf.py `
  --model-dir <path-to-Kandinsky-5> `
  --output-dir .\models\q8 `
  --type q8_0 `
  --dit <dit_path> `
  --qwen <qwen_path> `
  --clip <clip_path> `
  --vae3d <vae3d_path>

Use

This repo is intended for the native C/C++ runtime in kandinsky.cpp:

• code: https://github.com/AkaneTendo25/kandinsky.cpp

Use and redistribution should follow the terms of the original Kandinsky model release.