ZONOS3-Mini: local setup, licenses and browser limits

ZONOS3-Mini gives developers several ways to add local speech generation and voice cloning to an application, alongside a separately billed hosted API. The useful question is which runtime fits the device and which features are actually available there.

Zyphra announced the model on November 9, 2026. This guide reviews the public access paths and documentation as of November 12. It is not a hands-on audio-quality review: we inspected the sources and browser entry point, but did not generate or benchmark speech. Zyphra release announcement.

Weights, code and hosted access have different terms

The Hugging Face model card lists Apache 2.0 for the weights and 43 supported languages. It links the public GGUF and browser model files, plus 27 default voices grouped by language. The model repository was readable without signing in; Zyphra’s runtime README says downloading the public weights does not require a token.

The inference code uses the MIT license. That distinction matters when preparing a dependency inventory: the code license is not the weights license. Keep the notices for each component, and review third-party dependencies separately. Neither license gives a user permission to impersonate someone or use a recording without the necessary rights.

Zyphra’s model page lists hosted inference at $10 per million UTF-8 bytes. That is a text-byte rate, not a per-minute audio price. It also displays one concurrent request and 100 daily generations. Those published service limits should not be applied to a self-hosted runtime. Check the account’s current terms before building around hosted capacity.

For basic ASCII text, one character occupies one byte; other characters can use multiple bytes. Estimate API costs from the text’s UTF-8 byte length rather than assuming every language has the same character-to-cost ratio. Local use instead consumes the device’s storage, memory, processing time and power.

The browser is the shortest route, with a real download cost

The public Lite voice-cloning page offers a simpler entry point than the full studio. Its local engine uses WebGPU. In our browser inspection, the page loaded but reported that no WebGPU adapter was available, leaving generation disabled. That is a limitation of the browser used for this check, not evidence that the model fails on supported devices.

The WebGPU runtime documentation estimates about 1 GB for the initial backbone and decoder download. First use of an uploaded reference clip can load additional encoder components of roughly 2.8 GB, plus the speaker encoder. Speech editing adds more files. A no-install demo can therefore still be a substantial download.

The browser caches model files for later visits. Test both a first visit and a repeat visit when judging startup time, and check the target browser’s GPU support before promising an offline workflow. The documentation says initialization can fail when WebGPU is missing or adapter limits are too low.

Choose a local runtime around the machine

Zyphra’s runtime repository provides a Linux NVIDIA route, Apple Silicon MLX, C++/ggml and WebGPU. Its Python workspace pins Python 3.12 and uses uv. The NVIDIA setup specifies CUDA 12.8, downloads roughly 9 GB of PyTorch weights with the default decoder, and compiles kernels on first startup. The documented test machine is an H100, not a minimum consumer-GPU requirement.

The top-level installation example still points at an SSH checkout named ZONOS3, while the public release is ZONOS3-Mini. Start from the public repository’s own clone control and current backend instructions rather than assuming that every copied command or older path is ready for an unauthenticated setup.

For Apple Silicon, the MLX instructions install the zonos3-mlx workspace package, download the GGUF folder and start servers/mlx/scripts/serve.sh. The server exposes the studio and Lite interfaces on the local machine. Additional alignment and transcription workers are separate dependencies; automatic transcription is optional.

The ggml instructions offer native synthesis without Python at runtime. Building requires a suitable C++ toolchain, CMake and Rust for text normalization. The default quantized backbone is about 0.66 GB, but a working voice system also needs a decoder and speaker components. Downloading the entire documented GGUF folder is roughly 9.3 GB, so the backbone size alone understates storage needs.

Speech editing is not identical across backends

Basic synthesis, cloning, automatic transcription, pronunciation tools and editing should be treated as separate acceptance checks. A studio control being visible does not establish that the selected backend has its required model files.

The browser capability table excludes automatic speech recognition, text normalization and phonemization. Editing requires the clip’s transcript. Its alignment path handles clips up to 30 seconds, and edit inputs are capped at 150 seconds. Longer clips need the documented word timings or bounded spans. These limits make short, prepared material a more straightforward starting point than arbitrary recordings.

The native editing setup lists a separate edit backbone and adapter, reference encoders and a Python worker. Keep the edit adapter in its documented folder rather than mixing it into the ordinary synthesis configuration. Check the server’s capabilities response before relying on editing or transcription in an integration.

Read the size and speed claims carefully

Zyphra’s model page specifies 886 million total parameters and 100 million active parameters. The launch post and Hugging Face architecture outline list 800 million total. The active count describes sparse computation; it is not the full model’s memory footprint. Quote the source and measure the complete runtime rather than sizing hardware from the 100M label.

The launch reports browser time to first audio below 570 ms on a Pixel 9 and below 150 ms on an M3 MacBook Air. These are Zyphra’s measurements, not ByteForward results. First-audio latency also answers a different question from downloading the model, processing a reference voice or finishing a long narration.

For a useful evaluation, try the target language, names and numbers, a consented reference voice, long passages and repeated sessions on the intended device. Compare cold startup, sustained speed and memory use as well as listening quality. Our Phonon TTS deployment guide covers another local speech option, including its different licensing and voice-preparation constraints.

Featured image: editorial illustration created for ByteForward with Zyphra’s official wordmark from its press logo kit.