Zyphra's ZONOS3-Mini Clones a Voice With 100M Active Parameters and Starts Speaking in Under 150 ms on a MacBook Air

Zyphra has released ZONOS3-Mini, a compact text-to-speech model that clones a voice from a short reference clip and is small enough to run in real time in a browser or on a phone. The San Francisco lab published the model on November 9, 2026, with weights on Hugging Face under the Apache-2.0 license and runtimes on GitHub. It is also served through Zyphra’s own hosted inference on NVIDIA H200 GPUs.

ZONOS3-Mini is a sparse mixture-of-experts (MoE) model with 800 million total parameters, of which about 100 million are active for any given step. Zyphra says it trained the model on 112 NVIDIA H100 GPUs in two and a half days.

Shrinking the previous generation

The release is a deliberate step down in size from ZONOS2. Zyphra writes that ZONOS2 showed an MoE design could reduce the usual trade-off between voice quality and inference speed, but “did so at the cost of 8B parameters on a codec running at 86 Hz,” which limited where it could be deployed.

ZONOS3-Mini addresses that at several layers. The biggest change is a new audio codec, Zyphra Audio Codec 1 (ZAC-1), which replaces the Descript Audio Codec. Zyphra says DAC’s 86 Hz frame rate makes it hard to model and slows real-time generation, and that under similar data constraints a model trained on ZAC-1 showed a tenfold reduction in word error rate compared with one trained on DAC, with better acoustic similarity.

The backbone borrows techniques from recent large language models. Its sequence mixers use a 3:1 mix of Gated DeltaNet layers and global attention, which Zyphra says improves throughput at longer context lengths, along with latent MoE, Q balancing and attention residuals. For speaker identity, Zyphra swapped the Qwen3-TTS speaker encoder for the MECT-B2 speaker-verification model, which it says improves speaker similarity and makes voices easier to manipulate. After pretraining, the model went through direct preference optimization and then Group Relative Policy Optimization to improve robustness.

Where it runs

Zyphra ships runtimes for browsers through WebGPU, a native C++ runtime built on GGML, and server support through SGLang and vLLM. The company reports time-to-first-audio of under 570 ms in a browser on a Pixel 9 phone and under 150 ms on an M3 MacBook Air. It describes the target range as running “from top of the line server GPUs to multiple years old phones.”

The model supports two ways of conditioning on a speaker. One, trained from the start of pretraining, places a separate reference clip in context and produces more natural output without copying the reference’s prosody exactly. The other uses a direct voice representation of the target clip and was added in a short fine-tuning stage.

Steering accent, age and delivery

Much of the launch page is given over to control. Identity can be adjusted along age, gender presentation, accent and voice texture through speaker-embedding vectors, with dedicated task vectors for accents. Delivery, including pace, emotion, projection and liveliness, is controlled through activation steering. Zyphra’s samples show the same narrator voice shifted between General American, Southern English and Australian accents, aged from a target of 25 to 75, and moved from a soft, close delivery to calling across a room.

Control can also apply to individual words. Zyphra says it monitors learned alignment heads in attention during streaming generation, so a phrase inside a sentence can be whispered while the rest is spoken normally. A separate edit model, built on a VoiceCraft recipe, lets users change words in a sentence after generation with the same controls. Zyphra notes that its settings are generation targets, and that the controls “do not guarantee a specific emotion.”

A new benchmark, scored by Zyphra

Alongside the model, Zyphra released ZTTS2-Eval, a second version of its TTS benchmark. Its clean test set now covers 35 languages and its in-the-wild set 20 languages. Zyphra also built a Zyphra TTS Capability Index v1 using item response theory over 420 items, each combining a metric, a test set and a language. On that index, Zyphra says ZONOS3-Mini ranks first among open models, followed by Fish S2 Pro and MOSS TTS v1.5.

Those rankings come from Zyphra’s own benchmark and have not been independently reproduced. The launch page also offers side-by-side audio samples against Fish Audio, Qwen, Cartesia, ElevenLabs and OmniVoice, which Zyphra presents as qualitative comparisons rather than scored results.