Is ZONOS3-Mini state of the art?Pending benchmarks.Announced Nov 9, 2026; no independent results yet.
Zyphra's open text-to-speech model with zero-shot voice cloning and word-level control of emotion, pace and delivery. A sparse MoE with 800 million total parameters and 100 million active, licensed Apache 2.0. It runs on-device (browser WebGPU, phones, native GGML/MLX) and on servers (SGLang, vLLM), and is also served on Zyphra Inference / Zyphra Cloud.
The details so far
- Lab
- Zyphra
- Announced
- Nov 9, 2026
- API
- Live
- Open weights
- Published · Public, ungated Hugging Face repo Zyphra/ZONOS3-Mini (PyTorch, GGUF, WebGPU shards)
- Parameters
- 800M total / 100M active (sparse MoE)
- Type
- Zero-shot voice-cloning text-to-speech; 11-layer hybrid of Gated DeltaNet and global attention (3:1) with latent MoE (Zyphra blog and GitHub README)
- Codec
- Zyphra Audio Codec 1 (ZAC-1). GitHub README: the model predicts 10 × 4096 dual-codec audio codes at 15 Hz; a causal multi-resolution iSTFT decoder yields 44.1 kHz audio
- Voice control
- Identity (age, gender, accent, quality) via speaker embeddings and accent task vectors; delivery (pace, emotion, projection, liveliness) via activation steering, including word-level steering from alignment attention heads, plus a post-generation edit model
- Speaker conditioning
- Two modes: a clip in context (pretrained; more natural, does not copy prosody exactly) or a direct voice representation of the target clip (short finetune after pretraining). Speaker encoder is MECT-B2 (Zyphra blog)
- Training
- 112 NVIDIA H100s in 2.5 days, then DPO and GRPO post-training (Zyphra blog)
- Runtimes
- Day-0 SGLang (blog also lists vLLM); native CUDA, MLX, C++/GGML; in-browser WebGPU. 27 default voices ship in the Hugging Face repo
- Hosted access
- Zyphra Inference on NVIDIA H200s and Zyphra Cloud (zyphra.com / cloud.zyphra.com)
- License
- Model weights Apache License 2.0 (blog and Hugging Face card); inference runtime repository MIT (GitHub Zyphra/ZONOS3-Mini LICENSE)
- Code
- https://github.com/Zyphra/ZONOS3-Mini
- Related
- Separate from earlier Zyphra Zonos-v0.1, ZONOS2 and Zamba models; those are not this page
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: runs in real time with under 570 ms time-to-first audio in the browser on a Pixel 9 and under 150 ms on an M3 MacBook Air (Zyphra blog).
Lab claim: under similar data constraints, a model trained with ZAC-1 has a 10x lower word error rate than one trained with Descript Audio Codec, with better acoustic similarity (Zyphra blog).
Lab claim: self-reported on Zyphra's own ZTTS2-Eval / TTS Capability Index v1 (420 items), ZONOS3-Mini ranks first among open models, followed by Fish S2 Pro and MOSS TTS v1.5.
Sources
LABZONOS3-Mini blog (Zyphra)zyphra.com/our-work/zonos3-mini · Nov 9, 2026LABZyphra on X: ZONOS3-Minix.com/ZyphraAI/status/2119914895861076151 · Nov 9, 2026LABZyphra on X: ZONOS3-Mini thread (weights, blog, code)x.com/ZyphraAI/status/2119914911837061605 · Nov 9, 2026REPOZyphra/ZONOS3-Mini on Hugging Facehuggingface.co/Zyphra/ZONOS3-Mini · Nov 9, 2026REPOZyphra/ZONOS3-Mini on GitHubgithub.com/Zyphra/ZONOS3-MiniTracking continues until an independent board lists it.