Zyphra has released the weights for ZONOS3-Mini, a compact TTS model that performs zero-shot voice cloning from recordings up to 60 seconds long, designs new voices through age, accent, and timbre parameters, and allows for the adjustment of emotions and delivery down to individual words. The model synthesizes speech in real-time directly in the browser and on mobile devices without a GPU server, is distributed under the Apache 2.0 license, and, according to Zyphra, ranks first among open TTS models on the internal ZTTS2-Eval benchmark.

image
image
image

What happened

Zyphra published ZONOS3-Mini along with a nearly complete production pipeline. The architecture is an 11-layer MoE with 64 experts and top-2 routing: 800 million shared and approximately 100 million active parameters, of which 9 layers operate on Gated DeltaNet and 2 on global attention. Sound is encoded by the proprietary ZAC-1 audio codec at a rate of 15 frames per second with 44.1 kHz output. The company cites time-to-first-audio measurements: less than 570 ms in the browser on a Pixel 9, less than 150 ms on a MacBook Air M3, and an average of 80 ms on an H100 with 8 parallel requests; peak throughput on the H100 reached 457 seconds of audio per second. Weights are available on Hugging Face under Apache 2.0, the demo runs in the browser via WebGPU, and the GitHub repository contains runtimes for SGLang, MLX, ggml, and WebGPU.

Context

Speech synthesis with voice cloning recently required either closed APIs or significant server-side GPU resources, so the release of an open real-time class model shifts the balance of power: ElevenLabs is explicitly named as a quality comparison point in Zyphra's materials, although open TTS models rarely appeared in such comparisons on equal footing. The technical bet of the release is a hybrid MoE architecture with linear (delta) attention in most layers instead of heavy global attention, as well as a rare stream of codec tokens that reduces the length of the processed sequence. Due to the sparse activation of experts, the model utilizes only a portion of its parameters at any given time — this allows it to fit within real-time constraints on client devices. When reading conclusions about quality, it is important to note that Zyphra's public comparison is conducted on its own benchmark, not on an independent testbed.

Why this matters for the industry

For companies and developers, ZONOS3-Mini is a ready-made building block: Apache 2.0 allows embedding voice cloning into commercial products without licensing negotiations, and operation in the browser and on mobile removes the requirement to maintain GPU servers. The marginal cost of open real-time cloning approaches zero, which undermines the pricing moat of closed voice APIs and, if the claimed quality and latencies are independently confirmed, will increase pressure on their pricing. Value is shifting from the model itself to data, UX, and integrations: it becomes harder to differentiate a product when the basic component is free. Over the next six months, a wave of edge-TTS products based on hybrid linear attention architectures, independent comparative tests with open competitors and closed services, and niche products such as dubbing with frame-by-frame emotional direction and brand voices assembled by parameters are likely.

Why this matters for users

The model can be tried right now: the demo works in the browser via WebGPU without installation, weights are on Hugging Face, and runtimes for servers (SGLang), Apple Silicon (MLX), standard processors (ggml), and WebGPU are collected on GitHub. A clone requires approximately a one-minute voice recording; voice design through age, accent, and timbre allows obtaining a unique sound without a real voice actor. Emotions, tempo, and delivery can be adjusted down to individual words, opening up scenarios for dubbing with emotional direction, audio versions of texts, personalized voice notifications, and voice assistants, while private local voiceover on the device does not send data to the cloud. According to official information, the model supports 43 languages, including Russian, however, the quality of synthesis in Russian should be checked on your own texts.

What is currently unknown / limitations

Key latency and throughput figures are vendor-provided and tied to specific hardware: Pixel 9, MacBook Air M3, and H100; tail latencies and behavior under load peaks are not disclosed in the sources, and 80 ms is an average only for 8 parallel requests. The only qualitative comparison is the internal ZTTS2-Eval benchmark, owned by Zyphra itself: metrics, test data, and methodology are not described, objective indicators like WER, speaker similarity, or MOS were not mentioned, so "first place among open TTS" remains a vendor hypothesis until independent MOS or ABX evaluations. Russian support is officially stated, but there is no independent data on its quality yet — this should be expected from the community. Infrastructure requirements for server production are also not disclosed in the sources.

Sources

Author

Look at AI, editorial team