coolthor/Kandinsky-6.0-Pro-distill-5s-NVFP4

Kandinsky 6.0 Pro (distilled) — NVFP4

NVFP4 weight-only quantization of the Kandinsky 6.0 Video Pro, distilled transformer. Kandinsky 6.0 is an open (MIT) 29 B video model that generates a 5-second clip together with its 44 kHz audio: you write the spoken line into the prompt, and the voice and the lip movements come out of the same pass.

18.7 GiB instead of 56.2 GiB (60.3 GB), so the whole transformer fits on a 32 GB RTX 5090. On our 5090 it cut the time for a 5-second clip from 206.7 s to about 141 s (same settings, SageAttention on, model load included).

Same prompt, same seed (42). Left: original bf16. Right: this NVFP4 file. Audio is from the left clip. Both were generated on our RTX 5090.

📝 Full write-up: Testing Kandinsky 6.0 Pro's lip-sync on an RTX 5090 · 繁體中文版 🎬 Shorts: English · 中文

What's in the repo

File	What it is
Kandinsky-6.0-Pro-distill-5s-nvfp4.safetensors	The quantized transformer, ComfyUI native format
kandinsky6_nvfp4.yaml	The exact quantization config we used
assets/*.mp4	The sample clips on this page

Only the transformer is here. Get the text encoders, VAEs and audio models from the official ComfyUI setup guide: Qwen2.5-VL 7B, CLIP-L, the HunyuanVideo VAE, the MMAudio VAE and the BigVGAN vocoder.

How it was made

• Tool: comfy-quants, using kabachuha's kand6 branch (commit 4b96626), which adds the Kandinsky 6 model adapter.
• Quantized: every Linear layer in the visual, video-text and audio-text transformer blocks, 1,840 layers in total. They are stored as packed FP4 with FP8 (E4M3) block scales and an FP32 per-tensor scale.
• Kept in bf16: the first two and last two visual blocks (0, 1, 58, 59), following the LTX-2 NVFP4 recipe, plus norms, embeddings, time embeddings and the output head.
• The conversion itself took 59 seconds on the 5090. It reads the source one tensor at a time, but expect around 50 GB of RAM in use while it writes the file.

How to use it

1. ComfyUI 0.38.0 or newer, with the official kandinsky6 node installed.
2. Put the .safetensors file in models/diffusion_models/ and pick it in Load Diffusion Model (weight_dtype = default). ComfyUI loads NVFP4 natively; no node patches.
3. Sample the way you would sample the distilled checkpoint: 10 steps, CFG 1.0, audio VAE scaling 0.417.
4. Launch ComfyUI with --use-sage-attention (with sageattention installed). On our card, SageAttention alone cut the bf16 time from 277 s to 206.7 s.

Spoken lines go inside the video caption as <S>…<E>, for example She looks into the camera and says: <S>Welcome to AI Muninn.<E>. Describe the voice in the audio caption.

Our numbers (RTX 5090 32 GB, 61.6 GB RAM, Windows)

864×480, 121 frames at 24 fps, seed 42. The times include loading the model, and each is a single run.

Setup	Time per 5-second clip
bf16 (60.3 GB), no SageAttention	277 s
bf16 + SageAttention	206.7 s
NVFP4 (this repo) + SageAttention	141.5 s, 131.4 s, 151.5 s (three runs)

Speech, transcribed with Whisper large-v3-turbo:

Line we asked for	bf16	NVFP4
"Welcome to AI Muninn. Here's today's AI news."	Correct (it hears "Munnin")	Correct (it hears "Manin")
「歡迎來到 AI Muninn，今天帶你看最新的 AI 消息。」	Intelligible, 2 words wrong	Intelligible, 2 words wrong

The NVFP4 clip with the Chinese line. It is intelligible, but it sounds like a non-native speaker.

Caveats

1. Blackwell only for the speed. NVFP4 runs on tensor cores only on sm_120 and newer. Older cards can still load the file, but they expand the weights before each matmul, so all you gain is a smaller download.
2. Quantization changes the clip. A fixed seed does not give you the same clip faster. In our side-by-side the framing, outfit and lip timing stay almost identical and we saw no added noise, but the faces differ in small details. One prompt is not a quality benchmark.
3. English and Russian, not Chinese. The official report says the training prompts were Russian or English. Our Chinese line came out understandable but accented, in both bf16 and NVFP4.
4. The official ComfyUI node does not load the distilled Lite checkpoint, so we did not make an NVFP4 Lite.

Credits and license

The weights are Kandinsky 6.0 by Kandinsky Lab, released under the MIT license; this quantized copy keeps that license. Report: arXiv 2610.05608. Quantization by ai-muninn.

---

Made by Thor Lin. Full benchmarks, setup notes and the runs that failed are on ai-muninn.com. More quantized models: huggingface.co/coolthor.

Provenance

This section documents release 19c396245f94cb0e533f5efd9ebb427a6144d5e5, inspected on 2026-11-14. Where the exact upstream commit used at build time is unknown, it is listed as not recorded rather than backfilled with today's upstream HEAD.

Source lineage

Source repository	Revision used for this build	Role
kandinskylab/Kandinsky-6.0-Pro-distill-5s-Diffusers	not recorded	Immediate distilled BF16 input

Build details

Tools and versions: kabachuha/comfy-quants, kand6 branch, commit 4b96626 (short commit recorded in the card/blog). Full commit and exact invocation not recovered.

The model was quantized to weight-only NVFP4 using amax block scaling on in_features and nearest-even rounding. The conversion quantized 1,840 Linear layers across visual/video-text/audio-text blocks; visual blocks 0, 1, 58 and 59 remain BF16, as do norms and biases. The exact published config is kandinsky6_nvfp4.yaml. The upstream checkpoint was already distilled.

The existing records cover quantization and format conversion only; they don't mention any fine-tuning or merging by coolthor. That reflects the documentation, not an independent weight diff.

Exact build command: not recorded in the recovered material. The method description above is not a complete executable reproduction recipe.

Published conversion config: kandinsky6_nvfp4.yaml.

Weight file SHA-256 checksums

The values below are Hugging Face LFS object SHA-256 values for the pinned release above. File sizes and digests agree with pinned tree metadata. They identify full file contents, not LFS pointer text or Git blob IDs. Calibration imatrices are included when present.

File	Bytes	SHA-256
Kandinsky-6.0-Pro-distill-5s-nvfp4.safetensors	20094169424	dbed847427f149f90a5dfdc4bcd52a8d3b5f9378c8ad6ede2206ee0fb1e1a7e2

Verifying the files

Download the desired files from revision 19c396245f94cb0e533f5efd9ebb427a6144d5e5 and compare sha256sum <file> (Linux) or shasum -a 256 <file> (macOS) with the table. If cloning with Git LFS, hash the downloaded weight files, not the LFS pointer files. A matching hash confirms your files match this release; it does not prove the weights came only from the listed base model with no other changes.

Checking how the weights were derived requires the exact upstream commit, tool versions, recipe and any calibration or training data, then re-running the build and diffing the weights with a format-aware tool (or comparing file hashes where serialization is deterministic). Because of the gaps listed below, this build can't be reproduced exactly today. Quantized files are expected to differ numerically from BF16 sources, and a raw file hash cannot establish a base-to-quant weight diff. No quantization rerun or weight-diff audit was performed for this provenance review.

Known gaps: Base revision; Full converter commit and exact invocation.

Sources: existing model card and HF metadata.