ZONOS3-Mini explained: 100M-active open TTS, 43 languages

By LLMTest Lab · Nov 11, 2026 · 4 min read voiceopen-weightstext-to-speech
On this page

On this page

  1. What it is
  2. What is new
  3. The claims, and who made them
  4. Who should care
  5. What we will test

Not tested by us yet. Our test harness calls chat models over an API, and ZONOS3-Mini ships as open weights with runtimes you host yourself, so we have not listened to it ourselves. Everything below comes from Zyphra's model card on Hugging Face and its launch post. We'll update this post when we can run it on a fixed set of scripts.

ZONOS3-Mini is an open text-to-speech model from Zyphra that clones a voice from a short reference clip and can run in a browser tab. It has 800M parameters in total but only about 100M active per token, an Apache 2.0 licence for the weights, and support for 43 languages. If you need a self-hostable voice for an app and do not want a per-character bill, it is worth a listen. If you need a verified quality ranking, nobody outside Zyphra has published one yet.

ZONOS3-Mini
Maker Zyphra
Released 2026-11-09
Parameters 800M total, about 100M active (mixture of experts, 64 experts, top-2)
Licence Apache 2.0 (weights)
Output 44.1 kHz audio, up to 10 minutes per generation
Languages 43
Voice cloning Reference clip up to 60 s, or a single speaker embedding
Time to first audio 80 ms median at 8 concurrent requests on one H100 80GB (model card)
Training 112 H100 GPUs for 2.5 days (Zyphra)
Price Weights free; hosted pricing on Zyphra Cloud not published

What it is

The backbone is small: 11 layers, 9 of them Gated DeltaNet layers and 2 global attention layers, with a latent mixture of experts on top. That design is why 800M stored parameters behave like a 100M model at generation time. Audio is produced as tokens at 15 per second by a 44.1 kHz codec with 10 codebooks, then decoded by a causal iSTFT head. The model card also lists controls for loudness, bandwidth and speaking rate, and says emotion and pace can be steered down to individual words.

Voice cloning does not need a transcript. The card suggests a 5 to 10 second clip, and 27 default voices ship in the repository. There is also a "voice design" mode for shaping age, accent and texture without a reference recording.

What is new

Two things stand out. The first is the size-to-feature ratio: voice cloning, 43 languages and 44.1 kHz output from about 100M active parameters. The second is where it runs. Zyphra lists runtimes for SGLang and vLLM servers, an MLX server for Apple Silicon, a C++ ggml build for CPU, CUDA, Metal and Vulkan, and WebGPU for browsers. The weights are cached after the first visit, so the page works without an install or a server.

Zyphra reports time to first audio under 570 ms in the browser on a Pixel 9 and under 150 ms on an M3 MacBook Air. Those are the maker's numbers, and the model card does not state a real-time factor.

The claims, and who made them

Every quality claim here is Zyphra's. It scores systems on its own ZTTS2-Eval, which combines FLEURS-R read speech in 35 languages with VoxBlink2 in-the-wild recordings in 20 languages, and ranks ZONOS3-Mini first among open models on its own capability index, ahead of Fish S2 Pro and MOSS TTS v1.5. MarkTechPost reports 7.79 word error rate and 83.4 speaker similarity on the clean set, but that is Zyphra's evaluation too. Zyphra compares itself with ElevenLabs, Cartesia, Fish Audio and Qwen only in charts, and the launch post gives no absolute scores for those rivals in text.

Three caveats the model card is open about:

  • Text normalization covers only English, Mandarin, Japanese, Korean, German, Spanish, French, Italian and Portuguese. Other languages pass through as written.
  • Training data is not described.
  • The Hugging Face GGUF panel shows 11.3M parameters, which conflicts with the 800M and 100M figures. It may describe one file, such as the speaker component, but the card does not say.

Who should care

  • App builders who want a self-hosted voice. Apache 2.0 weights and a ggml build mean you can run it on your own hardware. Compare that to a hosted API, where cost scales with characters.
  • Anyone shipping voice on-device or in a browser. Few open voice-cloning models claim phone-class latency. Our write-up of Underdog Woof, a 32MB open TTS model on CPU covers the much smaller end of the same trade-off, and Alpie Seed does it for coding models.
  • Teams comparing open models for on-prem work. See which open LLMs actually run where for how we think about hardware limits.

Skip it if you need a published, independent ranking against ElevenLabs today, or if your target languages sit outside the nine with text normalization.

What we will test

We would read 20 short scripts in your target languages, including numbers, dates and product names, through ZONOS3-Mini and two hosted voices. We would score word error rate with a speech recognizer, then clone three voices from 10 second clips and check speaker similarity. Latency and cost per hour of audio go in the same table. Until then, treat every quality figure above as a vendor claim.

Related articles

Underdog Woof first look: a 32MB open TTS model on CPU
Underdog Woof is a 38M-parameter English text-to-speech model that runs on CPU in a 31.7 MB download. Specs, vendor-reported WER, license limits, who it suits.
Alpie Seed explained: a 9B Apache coding model for laptops
169Pi's Alpie Seed is a 9B Apache 2.0 model built on Qwen3.5-9B that runs offline. Specs, vendor benchmark claims, hardware needs and who should care.