Wildminder/VibeVoice-Quants

VibeVoice Quants

Quantized VibeVoice checkpoints for the ComfyUI-VibeVoice node pack.

Node repository  ·  Documentation  ·  Report an issue

These files are optional. The node downloads the official full-precision models on first run; point Load VibeVoice Model at a file from this repo to run quantized weights instead. No sidecar configs are needed — the node ships its own for every family here.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

❖ Files

File	Format	Size	Base model
VibeVoice-1.5B-bf16	bf16	5.4 GB	microsoft/VibeVoice-1.5B
VibeVoice-1.5B-fp8_e4m3	fp8	2.8 GB	microsoft/VibeVoice-1.5B
VibeVoice-1.5B-int8-row-convrot-gs64	int8	3.1 GB	microsoft/VibeVoice-1.5B
VibeVoice-1.5B-q8_0	GGUF	3.2 GB	microsoft/VibeVoice-1.5B
VibeVoice-7B-fp8_e4m3	fp8	9.5 GB	vibevoice/VibeVoice-7B
VibeVoice-ASR-HF-bf16	bf16	16.7 GB	microsoft/VibeVoice-ASR-HF
VibeVoice-Realtime-0.5B-bf16	bf16	2.0 GB	microsoft/VibeVoice-Realtime-0.5B
VibeVoice-Realtime-0.5B-int8_block-simple-heur	int8	1.1 GB	microsoft/VibeVoice-Realtime-0.5B

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

⌬ Formats

Format	Storage	Notes
bf16	full precision	Reference quality; matches the official checkpoints.
fp8	8-bit float, per-block scales	~1.9× smaller than bf16; weights stay packed in memory at run time.
int8	8-bit integer, rowwise/blockwise	The convrot variant stores rotated weights; simple-heur is blockwise.
GGUF	q8_0, llama.cpp-style	Single file; needs the gguf package (pip install gguf).

Quantization metadata is embedded in each checkpoint, so the node detects the format and the architecture from the file itself — no manual setup.

◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆◇◆

⇩ How to use

1. Download a file into ComfyUI/models/diffusion_models/.
2. Wire Load VibeVoice Model → external_model on the TTS or ASR node — connected, it overrides the standard model_name dropdown.
3. Keep config_name on Auto-detect.

Full setup, node inputs and VRAM notes in the node documentation.

§ License

MIT, matching the base models. VibeVoice is Microsoft's.