SSyncDev

ZONOS3-Mini

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

ZONOS3-Mini is Zyphra's open multilingual text-to-speech model: 800M MoE with ~100M active, zero-shot voice cloning in 43 languages and 80 ms first audio.

Key specifications

Specification score
25.1
Context window
32.77K
License
Apache-2.0 (weights); MIT (runtimes)

The specification score summarizes recorded features and limits. It is not a benchmark result.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Audio input

Produces

  • Audio output

Capabilities

AudioOpen weights

What is ZONOS3-Mini?

Under 570 ms to first audio in a phone's browser. Under 150 ms on a MacBook Air. About 80 ms on a server GPU. ZONOS3-Mini, released by Zyphra on 12 November 2026, is a text-to-speech model built to run wherever your users are, and it clones a voice from up to a minute of reference audio in 43 languages.

Weights are Apache-2.0; the runtimes and reference code are MIT.

Small active footprint

It stores about 800M parameters but activates roughly 100M per token. Eleven layers combine nine Gated DeltaNet blocks with two global attention blocks, feeding a 64-expert latent mixture-of-experts with top-2 routing. Audio is decoded through Zyphra's ZAC-1 codec at 15 Hz to 44.1 kHz output. The 32,768-position context covers about ten minutes of continuous speech.

The Hub lists 0.9B parameters against the 800M quoted in the release text. A rounding difference, most likely, but size memory for the larger figure. The release text also dates itself a day earlier than its publication date.

Where it runs

Runtime Notes
SGLang-Omni OpenAI-compatible streaming /v1/audio/speech; sgl-omni serve --model-path Zyphra/ZONOS3-Mini
MLX Apple Silicon
C++ / ggml CPU, CUDA, Metal and Vulkan with quantised GGUF weights
WebGPU In-browser Voice Studio

On a single H100 at concurrency 8 the median time to first audio is 80 ms, with peak throughput of 457x real time. Zyphra's figures.

The control surface

This is the reason to choose it over a plain voice model:

  • Acoustics: loudness, bandwidth, speaking rate.
  • Identity: age, gender presentation, vocal texture (soft, raspy, bright, deep) and accent steering.
  • Delivery: emotion, pace, projection, liveliness and whisper, applied to a whole line or to a chosen word span.
  • Editing: a separate checkpoint replaces a word or phrase inside existing speech, paired with an open word aligner.

Cloning works from an in-context reference of up to 60 seconds, or from a single speaker embedding.

Before you clone anyone

Voice cloning raises consent questions that are yours to answer, not the model's. Build your pipeline so that reference audio comes only from speakers who agreed, and log it.

A practical bake-off

Write 40 lines your product actually says, including numbers, names and a few emotional reads. Generate them in your top three languages on the runtime you would ship, browser included if relevant. Score naturalness with listeners and time to first audio on real devices. Compare with Voxtral TTS and Gemini 3.8 Flash TTS, or browse text-to-speech models.

How to evaluate ZONOS3-Mini

Compare the model on the work you need to complete.

Tasks and inputs

ZONOS3-Mini is listed for Audio, Text-to-speech. It accepts text, audio and produces audio. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 32.77K context. Compare pronunciation, voice consistency and time to the first playable audio. Check the service’s billing unit for the actual script.

Choosing a deployment

The weights are downloadable under Apache-2.0 (weights); MIT (runtimes). Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

43 languages, with zero-shot voice cloning from up to 60 seconds of reference audio.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

Alibaba

Qwen-Audio-3.1-TTS Plus

Qwen-Audio-3.1-TTS Plus is Alibaba's premium text-to-speech model for brand voices and dubbing, with fine-grained emotion tags and noise-robust voice cloning.

— context

Mistral AI

Voxtral TTS 3

Voxtral TTS 3 is Mistral's speech synthesis model for batch and streaming audio, with voice cloning, emotion tags and 35+ languages at $13 per 1M chars.

— context

fish-audio

Drama 3

Fish Audio Drama 3 is an expressive 35B MoE text-to-speech model steered by bracketed directions, covering 83 languages, with downloadable weights.

— context

heygen

HeyGen Voice

HeyGen Voice is HeyGen's default text-to-speech model: about 39 languages, public and cloned voices, streaming with character timestamps.

— context

heygen

HeyGen Voice Turbo

HeyGen Voice Turbo is the low-latency version of HeyGen's TTS model for public voices and instant clones, at $4 per 1M characters.

— context

Alibaba

Qwen-Audio-3.1-TTS Flash

Qwen-Audio-3.1-TTS Flash is Alibaba's streaming text-to-speech model for voice assistants, live chat and customer service, with system and cloned voices.

— context