Lab · 20-language cloning
MOSS-TTS-Nano: specs & requirements
MOSS-TTS-Nano is a 100-million-parameter autoregressive model from the OpenMOSS team (April 2026), and the only model here that ships an official browser-first ONNX stack — split prefill/decode graphs designed for onnxruntime-web. Voice cloning from a short reference recording is its primary workflow, it speaks 20 languages including Japanese and Korean, and it outputs native 48 kHz stereo — the highest fidelity in this lineup.
Lab tier — heavy/frontier model for enthusiast hardware
MOSS-TTS-Nano support is being wired into the tool. Its specs and hardware verdicts are below — meanwhile, the live models are ready to use, and the compatibility check already tells you whether your device could run MOSS-TTS-Nano.
Specifications
| Developer | OpenMOSS / MOSI.AI (Fudan) |
|---|---|
| Parameters | 100 million |
| Download size | 684 MB (fp32 (official ONNX export)) |
| Output | 48.0 kHz mono |
| Voices | 17 |
| Languages | Chinese, English, German, Spanish, French, Japanese, Italian, Hungarian, Korean, Russian, Persian, Arabic, Polish, Portuguese, Czech, Danish, Swedish, Greek, Turkish, Hebrew |
| Voice cloning | Yes |
| Streaming | Yes — audio starts before generation finishes |
| License | Apache-2.0 |
| Speed | Designed for real-time CPU inference; the official browser build streams audio as it decodes. |
| Runtime | WebAssembly (any CPU) |
Strengths
- +Official first-party browser deployment (onnxruntime-web) — not a community port
- +Voice cloning as the main workflow, across 20 languages
- +48 kHz stereo output; streaming-friendly split-graph export
Limitations
- −684 MB download (model + audio tokenizer), fp32-only export so far
- −Slower than the small models — expect several seconds of compute per sentence on CPU
- −9 curated built-in voices plus 8 licensed character references; several upstream presets are excluded under our voice-consent policy
Voices (19)
MOSS-TTS-Nano — questions
Is MOSS-TTS-Nano free to use?
Yes. The weights are Apache-2.0 and they run on your device.
What hardware does MOSS-TTS-Nano need?
Any modern browser — it runs on CPU via WebAssembly, no GPU needed. The download is 684 MB, cached after first use.
Is my text private when using MOSS-TTS-Nano here?
Yes. The model executes inside your browser tab — the text you synthesize is never transmitted to a server.
Can I use MOSS-TTS-Nano output commercially?
Review the current model, voice, and source-dataset terms before commercial use. The catalog lists Apache-2.0 for this model.