Zyphra has released ZONOS3-Mini, an open-source text-to-speech model with 800 million total parameters and approximately 100 million active parameters per token, distributed under the Apache 2.0 license. Published on November 12, 2026, the model generates 44.1 kHz audio at 15 Hz and supports voice cloning across 43 languages. The release targets developers building real-time voice assistants, localized content pipelines, and interactive media applications that require low-latency inference on consumer hardware.
"ZONOS3-Mini is Zyphra’s compact, open-source text-to-speech model, delivering natural reliable speech and high-fidelity voice cloning with just 100 million active parameters."
The model addresses a persistent bottleneck in generative audio: balancing high-fidelity synthesis with efficient compute. By combining a latent mixture-of-experts backbone with a causal multi-resolution iSTFT decoder, ZONOS3-Mini achieves a median time-to-first-audio of 80 milliseconds on an H100 GPU while maintaining competitive quality scores on standardized benchmarks. Its open weights and extensive runtime support lower the barrier for practitioners who previously relied on proprietary APIs for production-grade voice cloning.
Architecture and Tokenization
The model backbone consists of 11 layers with a model dimension of 1,024. It employs a hybrid sequence mixer pattern labeled gdn3_global1, which stacks 9 Gated DeltaNet (GDN) layers alongside 2 global attention layers. The attention layers utilize 16 heads with 32 dimensions and 4 KV heads, operating over a context window of 32,768 positions. Weights are stored in bf16 precision, while the decoder heads execute in fp32.
Text input is processed through an 8,192-entry SuperBPE tokenizer. For English, Mandarin, Japanese, Korean, German, Spanish, French, Italian, and Portuguese, the pipeline applies NeMo text normalization at inference. All other languages pass through the tokenizer as written. The autoregressive language model predicts 15 Hz dual-codec frames using Zyphra Audio Codec 1 (ZAC 1). Each frame contains 10 codebooks with 4,096 entries each. The first codebook is semantic and distilled from a Wav2Vec2-BERT teacher model, while the remaining nine capture acoustic details. A causal multi-resolution iSTFT head renders the frames to audio as they are generated, supporting up to 10 minutes of continuous output per generation.
Illustrative image: cliffsmith23 / Pixabay
Speaker conditioning, control rows, and SuperBPE text enter the hybrid GDN/attention MoE backbone, which predicts all ten codebooks of the next frame under a four-drawn delay pattern. As soon as a frame's last codebook is generated, the codec's causal iSTFT head renders that frame to audio. The routing mechanism uses 64 experts with top-2 routing activated from layer 1 onward, enabling sparse computation that keeps active parameters near 100 million per token.
Voice Cloning and Control Parameters
ZONOS3-Mini supports two distinct conditioning pathways for speaker identity. The first accepts a raw reference audio clip, which the codec encodes into a token prefix. This method supports clips up to 60 seconds and applies speaker classifier-free guidance (CFG). The second pathway uses a single 192-dimensional MECT-B2 speaker embedding (ownmect_s2), which bypasses the need for an audio reference entirely.
Beyond speaker identity, the model exposes three bucketed acoustic controls that allow developers to steer delivery down to individual words. Loudness is adjustable across 10 LUFS buckets, bandwidth (f_max) spans 10 frequency buckets, and speaking rate is controlled via 30 UTF-8 bytes per second buckets. Each control row includes an "unknown" index to handle out-of-distribution inputs. The bucket edge values are stored in params.json and can be queried programmatically. Developers can design characters by shaping age, accent, and vocal texture, or steer emotion and pace with fine-grained precision.
Benchmark Performance and Throughput
Quality evaluations utilize the ZTTS2-Eval benchmark (version 2), which aggregates read speech from FLEURS-R across 35 languages and in-the-wild recordings from VoxBlink2 across 20 languages. The reported quality metric combines macro speaker similarity and macro word error rate (WER) using the formula: ½ z(similarity) − ½ z(WER). ZONOS3-Mini was measured using the speaker-embedding prompt configuration. Full results and the ZTTS2 capability index are documented in the accompanying technical report.
Inference benchmarks were conducted on a single H100 80GB GPU. Under a load of 8 concurrent requests, the model achieved a median time-to-first-audio of 80 milliseconds. Peak throughput reached 457 seconds of audio per second at the system's optimal concurrency level. These figures position the model as a viable candidate for high-concurrency production environments where latency and throughput directly impact user experience. The quality-speed tradeoff demonstrates that sparse MoE routing and causal decoding can sustain real-time generation without sacrificing multi-language coverage or voice cloning fidelity.
Illustrative image: terydanphiri / Pixabay
Deployment Runtimes and Developer Access
Zyphra provides native runtimes that share a unified HTTP API and power the Zonos Voice Studio. The ecosystem includes a CUDA server for NVIDIA GPUs, an MLX server for Apple Silicon, a C++/ggml server supporting CPU, CUDA, Metal, or Vulkan, and a WebGPU build that executes entirely in the browser. The browser runtime downloads the model on first visit and caches it locally, requiring no external server.
For API-driven workflows, the model is served via SGLang-Omni behind an OpenAI-compatible /v1/audio/speech endpoint. The endpoint supports streaming 44.1 kHz PCM audio, voice cloning from reference clips without transcripts, and configurable sampling options. Developers can initialize a server with sgl-omni serve --model-path Zyphra/ZONOS3-Mini --port 8000 and generate audio via standard HTTP POST requests. The model is also compatible with llama.cpp, Ollama, Docker Model Runner, Lemonade, Unsloth Desktop, and Atomic Chat. Weights are hosted on Hugging Face, where the repository recorded 93,909 downloads in the last month.
Language Coverage and Text Normalization
The model supports 43 languages distributed across Western and Southern Europe, Northern Europe, Central and Eastern Europe, the Middle East and Caucasus, East Asia, Southeast Asia, and South Asia. Western and Southern European coverage includes English, French, Spanish, German, Italian, Portuguese, Dutch, and Greek. Northern European languages include Swedish, Danish, Norwegian, and Finnish. Central and Eastern European support covers Polish, Czech, Slovak, Hungarian, Romanian, Bulgarian, Russian, and Ukrainian. The Middle East and Caucasus region includes Turkish, Arabic, Hebrew, and Georgian. East Asia covers Mandarin Chinese, Japanese, and Korean. Southeast Asia includes Vietnamese, Indonesian, Malay, Thai, and Tagalog. South Asia supports Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, and Urdu.
Text normalization remains a constraint for non-English and non-Chinese/Japanese/Korean/German/Spanish/French/Italian/Portuguese inputs. These languages bypass NeMo normalization and pass through the tokenizer as written, which may affect pronunciation consistency for languages with complex orthography or tonal systems. Developers working with these languages should validate output quality against domain-specific use cases and monitor control bucket interpolation for prosody drift.
| Configuration Type | Max Duration | Codebook Count | Speed/Performance Impact |
|---|---|---|---|
full_mr Decoder | N/A | 10 | Default rendering quality |
keep2_plain Decoder | N/A | 10 | Approximately 3x faster decode |
| Raw Reference Clip | Up to 60 seconds | 10 | Requires audio upload and encoding |
| MECT-B2 Speaker Embedding | N/A | 10 | Bypasses audio reference entirely |
The release of ZONOS3-Mini shifts the center of gravity for open-weight speech synthesis toward hybrid state-space and attention architectures. The combination of latent MoE routing, dual-codec tokenization, and causal iSTFT decoding demonstrates that sub-100-millisecond time-to-first-audio is achievable without sacrificing multi-language coverage or voice cloning fidelity. Open questions remain regarding long-form consistency beyond the 10-minute generation limit and the exact impact of bucketed control interpolation on prosody. Developers integrating the model into production pipelines should verify control bucket boundaries against their target latency budgets and test speaker embedding stability across diverse acoustic environments. The technical report, authored by Gabriel Clark, Sofian Mejjoute, George Close, and Beren Millidge, provides full architectural diagrams, evaluation methodologies, and sampling configurations for practitioners requiring deeper implementation details.
Sources
Pneumetron
PNEUMETRON EDITORIAL TEAM
Rajini Ravindra holds an M.A. in History from Mysore University (KSOU). Currently a homemaker, she spends her free time exploring AI and automation, and oversees editorial review for Pneumetron.
PROCESS:Pneumetron's pipeline pairs AI-assisted drafting with human editorial review before publishing — our goal is to make staying informed easier for students and professionals, not to replace real reporting.
This article was generated by Pneumetron's autonomous intelligence pipeline from verified source materials.
Open Source Document at huggingface.co ↗



