**Zyphra released ZONOS3-Mini on November 9, 2026 — a sparse mixture-of-experts text-to-speech model with about 100 million active parameters, Apache 2.0 weights on Hugging Face, and day-zero inference paths for WebGPU browsers, Apple MLX, ggml, and SGLang-Omni.
If you build voice agents, the trade Zyphra is pushing is familiar from ZHA long-context work: ship quality near much larger models while keeping deploy surfaces small enough for edge and browser.
TL;DR
| Spec | Zyphra documentation |
|---|---|
| Architecture | Sparse MoE, 64 experts, top-2 routing |
| Params | ~800M total, ~100M active |
| License | Apache 2.0 weights |
| Clone modes | Reference audio or embedding-based voice design (age, accent, texture sliders) |
| Latency (vendor) | 80 ms TTFA on 1× H100 (8 concurrent); under 570 ms TTFA in Pixel 9 browser |
| Serve | SGLang-Omni, vLLM-class stacks, WebGPU studio, Zyphra Cloud on H200 |
What changed versus ZONOS2
Zyphra's post positions ZONOS2 as proof that MoE helps TTS quality, but at 8B parameters and an 86 Hz codec that limited deployment. ZONOS3-Mini cuts active compute and codec overhead so the model can sit on phones, laptops, and tab-local WebGPU without a datacenter hop.
Training footprint: 112 NVIDIA H100s for 2.5 days, per Zyphra — small by frontier standards, consistent with a focused audio release rather than a general LLM.
Controls builders actually use
The release page emphasizes controllable speech, not just cloning:
- Identity — zero-shot clone from audio or designed voices via sliders.
- Accent maps — steer British, Australian, and other targets from a base recording.
- Delivery — emotion, pace, whisper spans, and word-level edits via alignment monitoring during streaming plus a post-hoc edit model.
Zyphra ZONOS3-Mini open voice cloning TTS runs locally with WebGPU and SGLang servers
For game dialogue and adaptive tutors, word-level control matters more than leaderboard WER — you can fix one mis-stressed token without regenerating a paragraph.
How to run it this week
Hugging Face weights
Start at Zyphra/ZONOS3-Mini for GGUF shards, WebGPU bundles, and 27 default voices grouped by language in gguf/voices/.
Browser (no install)
Zyphra hosts Zonos Voice Studio and a Lite clone UI that run fully client-side on WebGPU. First load downloads weights; the browser caches them afterward.
NVIDIA server
SGLang-Omni documents:
sgl-omni serve --model-path Zyphra/ZONOS3-Mini --port 8000
Then POST to /v1/audio/speech like other OpenAI-compatible TTS gateways — useful if you already route Fish Audio or cloud APIs through one proxy.
Apple Silicon and CPU
Zyphra's GitHub runtimes share one HTTP API across MLX, CUDA server, and ggml CLI builds — pick the binary that matches your fleet.
Benchmarks: read the footnotes
Zyphra's comparison table on the release page pits ZONOS3-Mini against Fish Audio S2 Pro, MOSS-TTS v1.5, and Qwen3-TTS 1.7B on ZTTS2-Eval with Capability Index 66.2 for ZONOS3-Mini.
| Model (open) | Active params (reported) | ZTTS2 Capability (Zyphra) | License |
|---|---|---|---|
| ZONOS3-Mini | ~100M MoE | 66.2 | Apache 2.0 |
| Fish S2 Pro | Not disclosed | 61.7 | Fish research license |
| MOSS-TTS v1.5 | Not disclosed | 60.2 | Apache 2.0 |
| Qwen3-TTS 1.7B | Not disclosed | 37.7 | Apache 2.0 |
Vendor tables are starting points. Clone your own eval set: short prompts, noisy reference clips, and reference-clip vs embedding modes — Zyphra notes weaker scores on some reference-clip settings in third-party summaries.
Where it fits the open voice stack
explainx.ai has tracked a crowded open TTS month:
- Fish Audio Live 3 for duplex agents.
- Confucius4 streaming ASR for input stacks.
- Edge0 Audio8 for sub-1B cloning.
ZONOS3-Mini's niche is Apache weights + browser runtime + MoE efficiency — a credible default when you need on-device disclosure stories or want to avoid Fish's commercial license friction.
Pair with semantic caching on the LLM side if you run full voice agents: TTS savings mean little if the reasoning model still re-streams every turn.
Voice-agent stack recipe
A minimal local voice agent with ZONOS3-Mini in November 2026 might look like:
| Layer | Suggested component |
|---|---|
| ASR | Confucius4 streaming ASR or vendor cloud |
| Reasoning | Your existing Claude Code or Codex harness |
| TTS | ZONOS3-Mini via WebGPU or SGLang |
| Orchestration | Semantic caching on text side only |
Zyphra's 15 Hz codec frame rate (per comparisons on the release page) lowers audio bitrate versus older 86 Hz ZONOS2 stacks — helpful when you stream TTS over mobile networks during demos.
Licensing comparison for commercial apps
| Model | License gotcha |
|---|---|
| ZONOS3-Mini | Apache 2.0 — friendly for SaaS |
| Fish S2 Pro | Separate commercial license per model card |
| Qwen3-TTS | Apache 2.0 but heavier VRAM |
| ElevenLabs closed API | Terms of service on cloning consent |
If legal only approved Apache weights, ZONOS3-Mini clears the default bar without a secondary contract.
What people are asking
| Question | Short answer |
|---|---|
| GPU size for self-host? | Start with one consumer NVIDIA or Apple Silicon MLX; Zyphra targets real-time on M3 Air. |
| Browser-only production? | Possible for demos; production usually wants server ggml for logging and consent gates. |
| Multilingual? | Hugging Face card cites 27 default voices; Zyphra evaluated 35 languages in tables — verify your locale in studio before shipping. |
Safety and product notes
Open voice cloning carries fraud and impersonation risk. Apache weights mean no gatekeeper — your app must implement consent, watermarking, and rate limits. For regulated lines (health, finance), keep humans in the loop on outbound calls.
Zyphra also sells hosted inference on H200s at published per-byte pricing on its models page — useful for A/B against self-hosted ggml when GPU ops is not your core competency.
Hardware matrix (vendor claims)
| Environment | Reported TTFA | Notes |
|---|---|---|
| 1× H100, 8 concurrent | 80 ms median | Datacenter baseline |
| M3 MacBook Air | under 150 ms | MLX path |
| Pixel 9 browser WebGPU | under 570 ms | First visit downloads weights |
| Zyphra Cloud H200 | Vendor SLA | Matches enterprise buyers wanting no ops |
Treat mobile browser numbers as demo-grade until you test on your target devices — thermal throttling changes streaming policies.
Integration with SGLang-Omni
The SGLang cookbook documents OpenAI-compatible /v1/audio/speech — drop-in for stacks already fronting vLLM vs SGLang routers. Authentication still needs Hugging Face gating for staging weights; production should pin revision hashes on the model card.
When not to pick ZONOS3-Mini
| Use case | Maybe pick something else |
|---|---|
| Need 80+ languages with vendor support | Fish advertises broader locale coverage |
| Need sub-100ms phone on old Android without WebGPU | ggml CPU path may miss real-time |
| Need duplex speech-to-speech | Fish Live 3 targets full-duplex agents |
| Regulatory watermark mandate | Plan post-processing; open weights do not ship legal compliance |
MoE mechanics for TTS (why 100M active matters)
Sparse MoE routes each token through two of 64 experts in Zyphra's description. For TTS, that means compute scales with active parameters, not the 800M total footprint on disk. Shipping GGUF quants therefore targets the experts you actually activate, which is how Zyphra chases Fish-class quality without 8B-class VRAM.
If you already run MoE LLMs in production, the ops lesson transfers: profile expert load imbalance on long passages — audio novels hit different experts than UI prompts.
Observability for voice pipelines
Log TTFA, WER on a golden set, and GPU memory headroom per deployment tier. Zyphra's public numbers are marketing-grade; your production SLO should include barge-in behavior if you pair TTS with full-duplex stacks.
Fine-tuning and customization
Apache 2.0 weights allow LoRA and on-device adaptation subject to your own consent policies. Zyphra's word-level edit model (mentioned on the release page) is separate from base generation — pipeline builders can add a human approval step before TTS hits telephony trunks.
Download the Hugging Face model card before you bake ZONOS3-Mini into a customer pitch — check gguf/ versus web/ shards, confirm your target runtime, and run a ten-sentence golden set in both reference-clip and embedding modes. Zyphra's strongest marketing claims sit on embedding mode; your users may paste noisy phone audio.
Cost-wise, compare Zyphra Cloud per-byte pricing on the models page against self-hosted H100 hours plus engineer time — Apache weights make the break-even math public. For hackathon demos, start in the browser studio; for anything customer-facing, move to SGLang behind your own auth so you control rate limits and consent banners.
Voice cloning regulations are tightening in parallel — pair technical tests with consent capture in your app layer. ZONOS3-Mini makes cloning easy; compliance is still your problem whether you stream from WebGPU or a datacenter GPU.
Citation for papers: use the BibTeX block on the Hugging Face card (zyphra2026zonos3mini) if you benchmark against ZONOS3-Mini in public leaderboards — it keeps credit with the authors and makes reproduction easier for reviewers. Open the release-day commit on Hugging Face if you need exact revision hashes for pinned production deploys. Re-run latency tests after every driver or browser upgrade — WebGPU performance shifts with GPU firmware more than server CUDA does.
Related reading
- Zyphra ZHA long-context architecture
- Fish Audio Live 3 duplex agents
- Miso One real-time TTS guide
- Edge0 Audio8 compact cloning
- Confucius4 streaming for voice agents
- What is KV cache for inference
- What is llama.cpp for local models
Model specs, benchmark tables, and runtime URLs match Zyphra's November 9, 2026 materials.