Zyphra open-sources ZONOS3-Mini: 100M-active-param TTS running real-time in the browser
pmttyji · reddit · 2026-11-11
Zyphra released ZONOS3-Mini, an open-source text-to-speech model under Apache 2.0: 800M total parameters with 100M active per token (latent MoE, 64 experts, top-2), supporting 43 languages and running in real time on phones, laptops, and even in the browser via WebGPU.
It clones voices from an audio sample or designs character voices via age, accent and vocal texture, with fine-grained control over emotion, pace and delivery down to individual words, generating up to 10 minutes per run. The architecture uses a SuperBPE tokenizer, 9 Gated DeltaNet layers plus 2 global attention layers, autoregressive 15Hz dual-codec frames (1 semantic + 9 acoustic codebooks) and a causal multi-resolution iSTFT head rendering 44.1kHz audio. Official runtimes cover CUDA, MLX, C++/ggml and WebGPU, all sharing one HTTP API; a hosted version is on Zyphra Cloud.
Related event: Zyphra Open-Sources Zonos3-Mini: On-Device Real-Time Voice Cloning TTS(6 posts)→
More from Multimodal
- Tencent Hunyuan Hy4 API live on Atlas Cloud, iterates its own renders into a 3D world — TencentHunyuan · 2026-11-11
- Fable 5.2 recreates Star Wars Geonosis battle in 1.5 hours with one prompt — FinanceYF5 · 2026-11-11
- One prompt, 96 seconds: Fable 5.2 condenses 5000 years of writing history into 4K film — FinanceYF5 · 2026-11-11
- Fable 5.2 showcase: four fully generated films, zero live shots — FinanceYF5 · 2026-11-11
- Midjourney Motion-Blur B&W Portrait Recipe: --sref + --profile + --stylize 50 Shared — michaelrabone · 2026-11-11
- Hollowlight LoRA Brings Cinematic Dream-Realism Lighting to Image Models — MoonbearAIArt · 2026-11-11