Zonos-v0.1 — GGUF

GGUF weights for Zyphra/Zonos-v0.1-transformer, ready to run with zonos1.cpp — a standalone ggml/GGUF C++ port of Zonos-v0.1. The entire pipeline (espeak-ng phonemization → prefix conditioner → 1.6B dense transformer with classifier-free guidance → DAC-44 kHz vocoder, plus the ResNet293 speaker encoder for voice cloning) runs as native C++ linking only libggml + gguf: no Python, no PyTorch at inference time. CPU, CUDA, Metal, and Vulkan use the same files.

Zonos-v0.1 is Zyphra's open-weight text-to-speech model, trained on 200k+ hours of multilingual speech, with expressive conditioning (emotion vector, pitch, speaking rate, fmax) and high-fidelity voice cloning from a short reference clip. Primary languages: English, Japanese, Chinese, French, and German.

Files

File	Size	Description
`zonos1-f16.gguf`	3.25 GB	F16 backbone — lossless from the bf16 checkpoint
`zonos1-q8_0.gguf`	1.77 GB	Q8_0 backbone — recommended; effectively lossless
`zonos1-q4_k.gguf`	0.94 GB	Q4_K backbone — smallest
`dac.gguf`	254 MB	DAC-44 kHz decoder (codes → waveform)
`spk-encoder.gguf`	198 MB	ResNet293-SimAM-ASP speaker encoder (wav → 128-d embedding, for voice cloning)

All quants keep the prefix conditioner (including the speaker projection) at full precision and hold the output head + audio embeddings one tier above the bulk; only the transformer blocks take the target quant. Unlike its MoE sibling ZONOS2, this dense backbone tolerates 4-bit quantization gracefully.

Benchmarks

Build	Size	KLD ↓	Top-1 ↑
F16 (ref)	3.25 GB	—	—
Q8_0	1.77 GB	0.0004	99.2%
Q4_K	0.94 GB	0.024	96.8%

KLD (mean) and Top-1 measure how closely each quant tracks the F16 backbone's per-frame logits on teacher-forced generation traces, scored with the zonos1-perplexity tool in zonos1.cpp. Both quants produce clean, ASR-verified audio end-to-end; Q8_0 is the recommended default, Q4_K when memory is tight.

Quick Start

The zonos1.cpp release archives bundle a start-zonos1 launcher that downloads these files automatically and opens the web UI — no manual setup. Or fetch them yourself and run the tools directly:

zonos1-cli zonos1-q8_0.gguf --tts "Hello, world." out.wav \
    --dac dac.gguf --gpu --seed 1

Add --clone voice.mp3 (with spk-encoder.gguf next to the backbone) to clone a voice from a reference clip.

HTTP server

zonos1-server mirrors the reference Zonos API — low-latency streaming PCM, an OpenAI /v1/audio/speech endpoint, in-process voice cloning, continuous batching, and a browser UI:

zonos1-server zonos1-q8_0.gguf --dac dac.gguf --spk-encoder spk-encoder.gguf --gpu

(dac.gguf / spk-encoder.gguf are auto-detected when they sit next to the backbone.)

See the zonos1.cpp README for build instructions, quantization (quantize-cli), the desktop app, and the full CLI reference.

Zyphra/ZONOS1-GGUF