Zyphra Open-Sources Zonos3-Mini: On-Device Real-Time Voice Cloning TTS
Zyphra released the open-source TTS model Zonos3-Mini on November 10, focused on voice cloning and on-device real-time inference. The weights are available on Hugging Face under the Apache 2.0 license, and it gained SGLang-Omni support on launch day.
Confirmed
- Total parameters: 800M, with only about 80M–100M activated per token (accounts vary slightly; pmttyji describes it as an implicit MoE with 64 experts, top-2)
- Supports 43 languages, voice cloning, and word-level control (realmrfakename)
- Fully capable of running on-device: real-time inference works on phones, laptops, and even browsers (WebGPU); realmrfakename notes that every voice in the demo video came from TTS running in real time on a single phone
- SGLang-Omni support landed on launch day: first-audio-packet latency of just 57 milliseconds on a single H100 (as relayed by spikedoanz)
- Licensed under Apache 2.0, with weights on Hugging Face (victormustar)
Why it matters
- The very low activated parameter count means high-quality voice cloning no longer depends on cloud-scale compute, benefiting privacy and offline use cases
- Direct browser support (WebGPU) dramatically lowers the barrier to deployment and experimentation
- In hands-on testing, victormustar found it natural-sounding and quite impressive across a few trials, suggesting open-source TTS quality is rapidly closing in on commercial offerings
2026-11-10 ~ 2026-11-11 · 6 related posts
Primary sources
- Zyphra open-sources ZONOS3-Mini: 100M-active-param TTS running real-time in the browser — pmttyji ·
- Zyphra open-sources ZONOS3-Mini, an on-device TTS with voice cloning running in real time on phones — realmrfakename ·
- Zyphra open-sources ZONOS3-Mini: 800M-param voice-cloning TTS that runs in your browser — victormustar ·
- [source] Zyphra open-sources ZONOS3-Mini, an on-device TTS with voice cloning running in real time on phones — realmrfakename · 2026-11-10
- Zonos3-Mini Voice-Cloning TTS Hits SGLang-Omni: 57ms First Audio on One H100 — spikedoanz · 2026-11-10
- [source] Zyphra open-sources ZONOS3-Mini: 800M-param voice-cloning TTS that runs in your browser — victormustar · 2026-11-10
- ZONOS3-Mini TTS model: Apache 2.0, 800M params (100M active), HF demo live — victormustar · 2026-11-11
- Zyphra's ZONOS3-Mini TTS model with voice cloning trends on Hugging Face — Zyphra · 2026-11-11
1 near-duplicate retellings: pmttyji