Zyphra · Released, not on our provider
ZONOS3-Mini (2026): Zyphra’s open voice-cloning TTS model, its 43 languages, speed figures, hardware and license
ZONOS3-Mini is Zyphra’s open text-to-speech model with 100M active parameters out of 800M total, 43 languages, 44.1 kHz audio and zero-shot voice cloning, released on November 9, 2026 as Apache 2.0 weights, and this site cannot run it because it is a text chat, not a speech generator.
Use it now
Chat
This chat starts with Qwen3.8 Flash, not ZONOS3-Mini.
≈ 1 credit per reply · See all models
New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.
Start chatting — 5 free credits- Maker
- Zyphra
- Type
- Text
- Released
- 2026-11-09
- OpenRouter id
- None
Checked
ZONOS3-Mini: facts, with sources
- Zyphra’s blog post introducing ZONOS3-Mini is dated November 9, 2026, and the Hugging Face repository Zyphra/ZONOS3-Mini was created the same day. The technical report was submitted to arXiv on November 12, 2026 (revised November 13).Source: zyphra.com · checked 2026-11-19
- The blog calls ZONOS3-Mini a sparse mixture-of-experts text-to-speech model with 100 million active parameters and 800 million total parameters, licensed under Apache 2.0, trained on 112 NVIDIA H100 GPUs in 2.5 days. The technical report’s abstract gives the backbone as 886M parameters with 100M active per token; we could not reconcile 800M with 886M, so we quote both.Source: zyphra.com · checked 2026-11-19
- The model card’s architecture table lists an 11-layer backbone with model dimension 1,024 (9 Gated DeltaNet layers and 2 global attention layers), a 32,768-position context, an 8,192-entry SuperBPE text tokenizer, 15 Hz audio frames of 10 codebooks with 4,096 entries each, 44.1 kHz output, up to 10 minutes of audio per generation, bf16 weights, and a speaker prompt that is either a reference clip of up to 60 seconds or a single 192-dimensional speaker embedding.Source: huggingface.co · checked 2026-11-19
- Zyphra’s voice list for the default voices (27 clips, each with stored reference codes and a speaker embedding) has 10 English voices with styles such as Irish, British, energetic, anime and announcer, one French voice, and two voices each for German, Hindi, Italian, Korean, Mandarin Chinese, Portuguese, Spanish and Telugu. The voices are character voices named on the list, not real people.Source: huggingface.co · checked 2026-11-19
- The card says ZONOS3-Mini runs in real time on phones and laptops, including in a browser, and can clone a voice from an audio sample or design a character by shaping age, accent and vocal texture, steering emotion, pace and delivery down to individual words. Zyphra’s blog adds that these controls are targets that do not guarantee a specific emotion, and describes an edit model that changes sentences in an existing clip with the same controls.Source: zyphra.com · checked 2026-11-19
- The model card says the model weights are released under Apache 2.0 and states no other commercial-use limit. The GitHub README says the ZONOS3 runtime code is under the MIT License, lists third-party components in a NOTICE file, and notes that an opt-in phonemizer add-on uses espeak-ng, which is GPL-3.0.Source: huggingface.co · checked 2026-11-19
- The model card and GitHub README we read state no consent or acceptable-use rule for voice cloning; Zyphra’s site has Terms of Use and a Privacy Policy that we did not read. The consent note on this page is ours, not Zyphra’s.Source: huggingface.co · checked 2026-11-19
- The card says a hosted version is available on Zyphra Cloud; the price is not on the pages we read.Source: huggingface.co · checked 2026-11-19
- ZONOS3-Mini is not on OpenRouter’s model list (checked November 19, 2026), and this site is a text chat with language models, so it cannot generate speech with it. This site can chat with small text models such as Qwen3.8 Flash, Inkling 2 Small and Hy4 Flash.Source: openrouter.ai · checked 2026-11-19 · second-hand
Our table
ZONOS3-Mini speed and accuracy: Zyphra’s own figures
- The model card reports, for one H100 80GB, a median time to first audio of 80 ms with 8 concurrent requests and a peak throughput of 457 seconds of audio per second at the best concurrency, in a quality-versus-speed chart whose quality score combines speaker similarity and word error rate over 55 language subsets of Zyphra’s own ZTTS2-Eval (35 read-speech languages and 20 in-the-wild languages).Source: huggingface.co · checked 2026-11-19
- Zyphra’s blog says it runs in real time with under 570 ms to first audio in the browser on a Pixel 9 and under 150 ms on an M3 MacBook Air, and that on its Zyphra TTS Capability Index v1 ZONOS3-Mini ranks first among open models, followed by Fish S2 Pro and MOSS TTS v1.5. These are Zyphra’s own figures on its own benchmark, and we did not run them.Source: zyphra.com · checked 2026-11-19
- The technical report’s abstract says that on the new ZTTS2-Eval ZONOS3-Mini reaches a 6.71% macro-average word error rate over 55 language sets (35 clean and 20 in the wild), "competitive with larger and slower" systems, and that the weights, inference code and training code are released.Source: arxiv.org · checked 2026-11-19
Maker’s figures
ZONOS3-Mini latency, throughput, error rate and training cost as printed by Zyphra
People who search for ZONOS3-Mini want to know how fast and how good a 100M-active-parameter voice model can be. Zyphra’s answer is 80 ms to first audio on a data-centre H100 and under 150 ms on a laptop, with a 6.71% word error rate on its own 55-language benchmark. All of it comes from Zyphra’s documents and none is independently tested, so treat the ranking claim as a vendor claim. Other open speech models on this site are Kitten TTS 2 Mini, which is far smaller and has no cloning, and XTTS v2, an older cloning model whose license bars commercial use.
| Measure | Figure as printed | Condition |
|---|---|---|
| Median time to first audio | 80 ms | One H100 80GB, 8 concurrent requests |
| Peak throughput | 457 seconds of audio per second | One H100 80GB, at the best concurrency |
| Time to first audio in a browser | Under 570 ms | Pixel 9, real time |
| Time to first audio on a laptop | Under 150 ms | M3 MacBook Air |
| Macro-average word error rate | 6.71% | ZTTS2-Eval, 55 language sets (35 clean, 20 in the wild) |
| Rank on Zyphra’s capability index | First among open models, then Fish S2 Pro and MOSS TTS v1.5 | Zyphra TTS Capability Index v1, Zyphra’s own benchmark |
| Training | 112 NVIDIA H100 GPUs, 2.5 days | As stated in the blog post |
Our table
How to run ZONOS3-Mini: browser, SGLang, native servers and Zyphra Cloud, and what each needs
- The model card names three ways to run it: in a browser through WebGPU on your own GPU (Zonos Voice Studio and a lighter voice-cloning page, with the model downloaded on the first visit), through SGLang-Omni on NVIDIA GPUs behind an OpenAI-compatible /v1/audio/speech endpoint with streaming and cloning from a reference clip, and through Zyphra’s native runtimes: a CUDA server, an MLX server for Apple Silicon, a C++/ggml server and CLI for CPU, CUDA, Metal or Vulkan, and the in-browser build.Source: huggingface.co · checked 2026-11-19
- The GitHub README’s backend table lists a CUDA server for Linux x86_64 with an NVIDIA GPU (tested on an H100), an MLX server for Apple Silicon, the C++/ggml server and CLI, and the in-browser server. Its prerequisites are uv with Python 3.12 and, for the CUDA server, the CUDA 12.8 toolkit; the first CUDA start compiles kernels for several minutes, and the PyTorch weights come to about 9 GB with the default decoder.Source: github.com · checked 2026-11-19
- The ggml README says the default backbone for the native server is a quantized file of 0.66 GB, that the whole gguf folder of the Hugging Face repository is 9.3 GB with the full-size decoder, and that building it needs a Rust toolchain for the text normaliser.Source: github.com · checked 2026-11-19
Where it runs
The ways to run ZONOS3-Mini and what each needs
The cheapest way to try ZONOS3-Mini is the browser build, which needs only a browser with WebGPU; the heaviest is the CUDA server. Which one fits depends on whether you have an NVIDIA GPU, an Apple Silicon Mac or only a CPU. Zyphra prints no minimum RAM figure on the pages we read. If you only need a small model to write or summarise text, Qwen3.8 Flash and Inkling 2 Small run in the chat on this site.
| Route | What it is | What it needs |
|---|---|---|
| In your browser (WebGPU) | Zonos Voice Studio and a lighter cloning page; runs on your own GPU, no server | A browser with WebGPU; the first visit downloads the model and the browser caches it |
| SGLang-Omni server | OpenAI-compatible /v1/audio/speech endpoint with streaming and cloning from a reference clip | NVIDIA GPUs; install SGLang-Omni and start sgl-omni serve |
| Zyphra CUDA server | Zonos Voice Studio served locally on port 1919 | Linux x86_64, NVIDIA GPU (tested on an H100), CUDA 12.8, uv with Python 3.12, about 9 GB of PyTorch weights; the first start compiles kernels for several minutes |
| Apple Silicon (MLX) | MLX server that runs the GGUF files | An Apple Silicon Mac and the gguf folder from the Hugging Face repository |
| C++/ggml server and CLI | Native server for CPU, CUDA, Metal or Vulkan, no Python at runtime | A Rust toolchain to build; default backbone 0.66 GB, whole gguf folder 9.3 GB |
| Zyphra Cloud (hosted) | A hosted version of ZONOS3-Mini run by Zyphra | An account with Zyphra; the price is not on the pages we read |
Our table
ZONOS3-Mini languages: the 43 languages by region
- The card says ZONOS3-Mini supports 43 languages and groups them by region: Western and Southern Europe, Northern Europe, Central and Eastern Europe, the Middle East and Caucasus, East Asia (Mandarin Chinese, Japanese, Korean), Southeast Asia and South Asia. Text is normalised with NeMo-based rules for English, Mandarin, Japanese, Korean, German, Spanish, French, Italian and Portuguese; other languages pass through as written.Source: huggingface.co · checked 2026-11-19
Languages
Languages ZONOS3-Mini supports, grouped as on the model card
ZONOS3-Mini covers many more languages than the small Kitten models, including eleven South Asian languages and Arabic, Hebrew and Georgian. The card says text is normalised with NeMo-based rules for nine of them and passed through as written for the rest, so numbers and dates in the other languages may be read less well; Zyphra gives no per-language quality table on the card.
| Region | Languages | Which |
|---|---|---|
| Western and Southern Europe | 8 | English, French, Spanish, German, Italian, Portuguese, Dutch, Greek |
| Northern Europe | 4 | Swedish, Danish, Norwegian, Finnish |
| Central and Eastern Europe | 8 | Polish, Czech, Slovak, Hungarian, Romanian, Bulgarian, Russian, Ukrainian |
| Middle East and Caucasus | 4 | Turkish, Arabic, Hebrew, Georgian |
| East Asia | 3 | Mandarin Chinese, Japanese, Korean |
| Southeast Asia | 5 | Vietnamese, Indonesian, Malay, Thai, Tagalog |
| South Asia | 11 | Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Urdu |
Get told when it is available
Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.
Nearest available
Alternatives to ZONOS3-Mini you can use now
Qwen3.8 Flash
Alibaba$0.15 / $0.47 (per million tokens)
- Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
- Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13
Inkling 2 Small
Thinking Machines Lab$0.30 / $1 (per million tokens)
- Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, and made it available on its Tinker platform.Source: thinkingmachines.ai · checked 2026-11-14
- A mixture-of-experts model with 594B total parameters and 17B active, with image input and variable effort.Source: thinkingmachines.ai · checked 2026-11-14
- Text and image input, text output; context window up to 1M tokens; each token is routed to 8 of 384 experts.Source: huggingface.co · checked 2026-11-14
Hy4 Flash
Tencent$0.04 / $0.14 (per million tokens)
- Tencent’s Hy4 flash announcement, dated November 16, 2026, presents it as the newest member of the Hy model family. Tencent recommends it for most everyday tasks (deep research, document processing, PowerPoint generation, simple front-end work, computer use) and points complex tasks such as long-horizon software engineering to Hy4. Tencent Cloud’s API guide calls it "a leaner, more efficient everyday workhorse".Source: tencent.com · checked 2026-11-17
- The Tencent announcement is dated November 16, 2026, and Tencent Cloud’s model, pricing and API pages for hy4-flash were last updated the same day. The announcement says only that Hy4 flash is available now; it gives no separate release date or version name.Source: tencent.com · checked 2026-11-17
- Tencent says Hy4 flash has 224B total parameters (including 64B Engram parameters), 12B active parameters, a context of up to 1M tokens and multimodal input; the announcement does not state a maximum output or the output types. It says the KV cache is 1/6 the size of Hy4’s and that decoding exceeds 150 tokens per second. These are the maker’s own figures, not tested by us.Source: tencent.com · checked 2026-11-17
FAQ
ZONOS3-Mini FAQ
What is ZONOS3-Mini?
ZONOS3-Mini is a text-to-speech model from Zyphra that clones a voice from an audio sample and lets you steer age, accent, emotion and pace. Zyphra’s blog dated November 9, 2026 describes a mixture-of-experts model with 100 million active and 800 million total parameters, 44.1 kHz audio and 43 languages. It runs on a server GPU, a Mac, a CPU build or in a browser.
Is ZONOS3-Mini open source, and can I use it commercially?
The model card says the weights are released under Apache 2.0 and states no other commercial-use limit, and the GitHub README puts the runtime code under the MIT License. Optional parts have their own terms, for example an opt-in phonemizer add-on that uses espeak-ng under GPL-3.0. Read the license files in the repositories you use.
What hardware does ZONOS3-Mini need?
It depends on the route. Zyphra’s CUDA server targets Linux with an NVIDIA GPU (tested on an H100) and about 9 GB of PyTorch weights; the C++/ggml build runs on CPU, CUDA, Metal or Vulkan with a 0.66 GB default backbone; and a browser build runs on your own GPU through WebGPU. Zyphra prints no minimum RAM figure on the pages we read.
How fast is ZONOS3-Mini?
By Zyphra’s own figures, which we did not run, the median time to first audio is 80 ms on one H100 with 8 concurrent requests, under 150 ms on an M3 MacBook Air and under 570 ms in a browser on a Pixel 9. Peak throughput is 457 seconds of audio per second on the H100. Its macro-average word error rate on Zyphra’s own benchmark is 6.71% across 55 language sets.
Which languages and voices does it have?
The card lists 43 languages, from English, French and Mandarin Chinese to eleven South Asian languages, Arabic, Hebrew and Georgian. There are 27 default voices (10 English and two each for most other listed languages), and you can also supply your own reference clip or a speaker embedding.
Can ZONOS3-Mini clone a voice, and what about consent?
Yes: it clones from a reference clip of up to 60 seconds or from a stored speaker embedding. Use only recordings of people who have agreed, do not present generated speech as a real person, and check the rules where you live; Zyphra’s card and README state no consent rule, and this note is ours.
Can I generate speech with ZONOS3-Mini here?
No. This site is a text chat with language models, so it cannot produce audio with ZONOS3-Mini or any other speech model, and the model is not on OpenRouter’s list. For text, a typical reply to Inkling 2 Small costs about 1 credit on this site (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Qwen3.8 Flash is another small model.
Is this the official Zyphra website?
No. This page is unofficial and not affiliated with or endorsed by Zyphra. Facts link to the pages we read.