Zyphra releases ZONOS3-Mini, an open text-to-speech model

Zyphra says training used 112 NVIDIA H100s over 2.5 days, with changes in the codec, the architecture, the speaker encoder and post-training[1].

The card describes an 11-layer backbone of dimension 1,024, a latent mixture of experts with 64 experts and top-2 routing from layer 1, and a context of 32,768 positions. The announcement calls the sequence mixers a 3:1 hybrid of Gated DeltaNet and global attention. The card counts nine Gated DeltaNet layers and two global attention layers. The model page lists the context as 32K and the quantisation as BF16[1][2][3].

NeMo normalises English, Mandarin, Japanese, Korean, German, Spanish, French, Italian and Portuguese. Other languages pass through as written. An 8,192-entry SuperBPE tokenizer feeds 15 Hz frames, ten codebooks of 4,096 entries, rendered to 44.1 kHz by a causal iSTFT head, for up to 10 minutes. The weights are bf16[2].

Zyphra calls the codec Zyphra Audio Codec 1 (ZAC-1). Under similar data constraints, Zyphra says a model trained with it has a word error rate 10 times lower than one trained with the Descript Audio Codec, and better acoustic similarity. Against S1-DAC and MOSS v2, Zyphra says it is comparable at better compression, or better at similar compression[1]. Voices use the MECT-B2 model rather than the Qwen3-TTS encoder. A clip kept in context from pretraining sounds more natural, Zyphra says, and does not copy prosody exactly. A later fine-tune uses a direct representation of the target. The card allows a raw reference of up to 60 seconds, or one 192-dimensional embedding[1][2]. Post-training is direct preference optimisation, then group relative policy optimisation[1].

The card lists 43 languages: English, French, Spanish, German, Italian, Portuguese, Dutch, Greek, Swedish, Danish, Norwegian, Finnish, Polish, Czech, Slovak, Hungarian, Romanian, Bulgarian, Russian, Ukrainian, Turkish, Arabic, Hebrew, Georgian, Mandarin Chinese, Japanese, Korean, Vietnamese, Indonesian, Malay, Thai, Tagalog, Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia and Urdu[2].

Zyphra says the server steers identity, accent, quality, mood and delivery. Word-level control, while audio streams, uses alignment heads. An edit model based on a VoiceCraft recipe applies the same controls after generation[1]. Age, gender, accent and quality use speaker embeddings and accent task vectors. Pace, emotion, projection and liveliness use activation steering. Loudness, bandwidth and speaking rate are bucketed on the card[1][2].

Zyphra calls these controls targets, not guaranteed traits or measured ages[1].

The figures are Zyphra's, on ZTTS2-Eval. A clean set covers 35 languages and an in-the-wild set covers 20. The card names FLEURS-R and VoxBlink2, 55 subsets. Qwen3-ASR transcribes them, with Whisper Large v3 added for lower-resource languages and for English, where Zyphra says that choice advantages Qwen models[1][2].

On Zyphra's TTS Capability Index v1, scored over 420 items, Zyphra ranks ZONOS3-Mini first among open models, followed by Fish S2 Pro and MOSS TTS v1.5[1]. Clips also sit beside Fish Audio, Qwen, Cartesia, ElevenLabs and OmniVoice. Zyphra prints no number for them and says listening is the better test[1].

On one H100 with 80 GB, Zyphra's quality figure is half a z-score of macro speaker similarity minus half a z-score of macro word error rate, using a speaker embedding. Zyphra reports 80 milliseconds median time to first audio at eight concurrent requests, and 457 seconds of audio per second at peak[2]. In a browser, Zyphra says that time is under 570 milliseconds on a Pixel 9 and under 150 milliseconds on an M3 MacBook Air[1].

The weights are under Apache 2.0[1][2][3]. The runtime repository the card links is under the MIT licence: CUDA on Mini-SGLang, MLX on Apple Silicon, C++ and GGML for CPU, CUDA, Metal or Vulkan, and WebGPU in the browser, cached after the first visit. The weight files include 27 default voices[2][4]. The card's server quick start is SGLang-Omni, which can clone from a clip with no transcript[2]. The announcement also names vLLM, phones several years old, and Zyphra Inference on NVIDIA H200s[1]. The model page prices zyphra/zonos3-mini at $10.00 per 1 million UTF-8 bytes, with one concurrent request, 100 generations a day, unlimited input characters, a 60-second raw clone and a 120-second embedding clone[3].

None of the four sources describes a consent check, a watermark, or a ban on cloning a given voice[1][2][3][4]. The README says the CUDA speech-edit adapter is in a private repository, while GGML and MLX load public GGUF edit files[4]. The announcement still calls the edit model released[1].

Zyphra's latency and open-model rank sit beside a voice cloner it says a browser can cache. None of the four sources says the person in the sample has to agree.