ZONOS2-GGUF

GGUF weights for [**Zyphra/ZONOS2**](https://huggingface.co/Zyphra/ZONOS2), ready to run with [**zonos2.cpp**](https://github.com/Zyphra/zonos2.cpp) — a standalone ggml/GGUF C++ port of ZONOS2. The **entire pipeline** (ECAPA speaker encoder → MoE backbone → DAC vocoder) runs as native C++ linking only `libggml` + `gguf`: **no Python, no PyTorch, no CUDA-only kernels at inference time.** CPU and CUDA use the same files. ZONOS2 is Zyphra's latest text-to-speech model, trained on more than 6 million hours of varied multilingual speech, delivering expressiveness and quality on par with—or even surpassing—top TTS providers at low latency with MoE. ZONOS2 excels at high-fidelity and naturalistic voice cloning. **For model details and speech samples, check out our [blog](https://www.zyphra.com/our-work/zonos2).** **A hosted version is available at [cloud.zyphra.com/audio-playground](https://cloud.zyphra.com/audio-playground).** Language support is as follows. All quants keep the **spine (attention + router + embeddings + speaker projection) at F16** and quantize only the MoE expert matrices — the layout that holds quality far below the usual quant floor. The full-precision spine is the single biggest quality lever; pick the expert precision (Q8 → Q4) that fits your VRAM. - **KLD** (mean) and **Top-1** measure how closely each quant tracks the F16 backbone's per-frame logits, scored with the `zonos2-perplexity` tool in zonos2.cpp. - **WER** (Qwen3-ASR word-error rate), **SpkSim** (clone speaker similarity), and **UTMOS** (predicted MOS) are end-to-end on the [`Zyphra/ZTT1-Eval`](https://github.com/Zyphra/ZTT1-Eval) Clean English set; lower WER and higher SpkSim/UTMOS are better. Although the logit metrics (KLD, Top-1) degrade steadily as the experts shrink, the **audio quality holds nearly flat down to Q4_K** — WER, speaker similarity, and UTMOS stay within eval noise of F16. The F16 spine keeps the model on-distribution, so the smaller quants spend their error budget on inaudible logit jitter rather than audible artifacts. **Q8_0 is the recommended default** (effectively lossless); Q4_K is a strong choice when VRAM is tight. Build [zonos2.cpp](https://github.com/Zyphra/zonos2.cpp) (CPU, or `-DGGML_CUDA=ON` for NVIDIA), then turn text into a waveform with one command: zonos2-cli zonos2-q8_0.gguf --tts "Hello, world." out.wav \ --dac dac.gguf --gpu --seed 1 Add `--spk voice.mp3` (with `spk-encoder.gguf`) to clone a voice from a reference clip. `zonos2-server` mirrors the reference FastAPI — low-latency streaming PCM, an OpenAI `/v1/audio/speech` endpoint, in-process voice cloning, and a browser UI: zonos2-server zonos2-q8_0.gguf --dac dac.gguf --spk-encoder spk-encoder.gguf --gpu See the [zonos2.cpp README](https://github.com/Zyphra/zonos2.cpp) for build instructions, quantization (`quantize-cli`), batching, and the full CLI reference. If you find this model useful in an academic context please cite as: @misc{zyphra2025zonos, title = {Zonos V2 Technical Report}, author = {Gabriel Clark, Sofian Mejjoute, Mohamed Osman, George Close, Beren Millidge}, year = {2026}, }