ZONOS3-Mini: 100M-Parameter TTS That Runs in a Browser
Zyphra has released ZONOS3-Mini, an open-source text-to-speech model designed for one specific capability that has been out of reach until now: real-time speech generation entirely on the device in front of you. The weights ship under Apache 2.0, and the company says the model runs in a browser tab, on a laptop, or on a server GPU.
The specifications explain the ambition. ZONOS3-Mini is a sparse Mixture of Experts with 800 million total parameters and about 100 million active per token, routed across 64 experts with top-2 selection. It supports zero-shot voice cloning from a reference clip, word-level control over how a sentence is delivered, and generations up to 10 minutes of audio. The 32,768-position context is modest by language-model standards and generous for speech.
Small, fast and open is a combination that changes who can build with speech, and the deployment list is the proof: a WebGPU playground that runs fully in the browser, an NVIDIA server runtime, native builds for Apple Silicon and CPU through ggml, and Zyphra's own cloud.
What ZONOS3-Mini is, and what came before
ZONOS3-Mini follows ZONOS2, Zyphra's 8 billion parameter model with an 86 Hz codec. ZONOS2 produced quality speech, and its size limited where it could run, which is the problem ZONOS3-Mini exists to solve.
The training budget is notable for what it says about efficiency. Zyphra trained the model on 112 NVIDIA H100s in 2.5 days. That is a small run by frontier standards, and it produced a model that Zyphra's own evaluation ranks first among open models on its Capability Index.
The lineage also explains the design choices. Zyphra has been building toward on-device speech for two years, and each generation has moved the same direction: better codecs, smaller active footprints, more deployment targets. The company's hierarchical attention research pushed on the context side, and ZONOS3-Mini pushes on the deployment side.
The three layers ZONOS3-Mini rebuilt
The announcement describes changes at three layers of the stack, and each one carries weight. Reporting on the release came from MarkTechPost, with Zyphra's own announcement as the primary source.
The first is the codec. Zyphra trained Zyphra Audio Codec 1, or ZAC-1, to replace the Descript Audio Codec used previously. ZAC-1 runs at 15 Hz with 10 codebooks of 4,096 entries each, one semantic and nine acoustic. Zyphra reports a 10 times word error rate reduction versus a DAC-trained model under similar data constraints. A causal iSTFT head renders each frame to 44.1 kHz audio as soon as it is generated, which is what makes streaming speech possible without waiting for a whole generation to finish.
The second is the backbone. The model uses 11 layers with a model dimension of 1,024, and sequence mixing follows a 3:1 ratio of Gated DeltaNet layers to global attention layers, nine and two respectively. Latent MoE, Q balancing and attention residuals round out the architecture. The sparse design is what keeps active parameters at 100 million while total parameters sit at 800 million, and that ratio is the difference between a browser runtime and a datacentre requirement.
The third is the speaker encoder. Voice representation moved from the Qwen3-TTS speaker encoder to MECT-B2, and the model accepts either a raw reference clip of up to 60 seconds or a single 192-dimensional speaker embedding. Post-training used DPO followed by GRPO to improve robustness.
For developers, the three layers add up to a practical conclusion. The codec determines how much compute each frame of audio costs, the backbone determines how fast those frames generate, and the speaker encoder determines how easily a voice can be specified. Rebuilding all three at once is how a model ends up small enough for a browser without losing the cloning feature that makes it useful, and it is what separates ZONOS3-Mini from smaller models that simply dropped capabilities to fit.
Why ZONOS3-Mini's 100M active parameters matter
The number to watch in this release is not the 800 million total. It is the 100 million active.
Active parameters are what each generated token actually costs in compute. A dense 8 billion parameter model spends all 8 billion on every token, which is why ZONOS2 needed serious hardware. A sparse model like ZONOS3-Mini spends a tenth of that per token, which is why it runs in a browser. The MoE design trades memory for speed, holding a larger model in storage while activating only a small slice per step, and speech generation with its tight latency budgets is exactly the workload where that trade pays.
The expert structure also fits speech oddly well. Speech synthesis draws on a mix of general skills, prosody, phoneme timing, acoustic texture, that apply to every utterance, plus specialised patterns that vary by speaker, language and emotion. Routing tokens to a small subset of experts gives the model a way to specialise per utterance without loading specialisation into every computation. Whether that intuition holds at scale is exactly what the benchmark is testing, and so far the answer on Zyphra's index is yes.
The pattern is now visible across the local-AI stack. Models are being redesigned around what modest hardware can activate per token, whether through sparsity, through attention architectures that cut context cost, or through compression that shrinks the weights themselves. ZONOS3-Mini applies the same logic to speech, where the problem is harder because the output must arrive continuously and on time.
For developers, the consequence is concrete. Speech interfaces, accessibility tools, real-time translation and voice agents no longer require a GPU server to sound good. A browser page can carry a cloned, steerable voice, and the weights to build it are Apache 2.0.
How ZONOS3-Mini performs on benchmarks
Zyphra evaluated ZONOS3-Mini on its new ZTTS2-Eval benchmark, covering 35 languages of clean read speech and 20 languages of in-the-wild recordings. A Capability Index built with item response theory over 420 items produces the headline number.
In speaker-embedding mode, ZONOS3-Mini scores 66.2 on that index. Fish Audio S2 Pro scores 61.7, MOSS-TTS v1.5 scores 60.2, and Qwen3-TTS 1.7B scores 37.7. On the clean set, ZONOS3-Mini posts a 7.79 word error rate and 83.4 speaker similarity, the best of the compared models on both. Those accuracy numbers matter for real deployments, because a speech system that mispronounces or mistranscribes is unusable no matter how good it sounds.
Every one of those numbers comes with an asterisk that Zyphra itself provides: Zyphra built and ran the benchmark. Independent verification does not exist yet, and the comparisons across models mix different GPUs and setups that the company explicitly labels as not directly comparable.
Two caveats are more important than the leaderboard. Reference-clip cloning mode scores only 44.7 on the index, far below the 66.2 of embedding mode, which means the convenient workflow, dropping in a short audio sample, is the weaker one, and the embedding workflow requires tooling around the model. And the clean-set MSR-UTMOS quality score of 3.07 trails Fish S2 Pro at 3.16 and Qwen3-TTS 1.7B at 3.54, meaning at least one competitor's audio quality is rated higher even as ZONOS3-Mini leads on accuracy and similarity.
A third caveat sits in the benchmark's design. The Capability Index measures the skills the benchmark's authors chose to measure, and the authors are the model's creators. The set covers clean read speech and in-the-wild recordings across 55 languages, which is a broad sweep, but emotional expressiveness, long-form narration and code-switching are not the centre of it. Teams with specific audio requirements should test on their own material before trusting the index.
An honest reading is that ZONOS3-Mini leads the open field on capability-per-parameter and on deployment flexibility, while the top of the quality leaderboard remains contested.
Word-level voice control
The most distinctive feature is control over how individual words are delivered, not just how they sound.
Identity controls cover age, gender, accent and voice texture through speaker embedding vectors and accent task vectors. Delivery controls cover pace, emotion, projection and liveliness through activation steering. Learned alignment heads let that steering apply to specific words during streaming, which means a developer can mark one word in a sentence to be delivered emphatically, quietly or quickly while the rest of the sentence flows normally.
A separate edit model, based on a VoiceCraft recipe, can modify sentences after generation, which turns a rendered clip into something closer to an editable document. For anyone who has generated a long speech clip and heard one wrong emphasis near the end, post-generation editing is not a gimmick, it is the difference between usable and regenerating.
Word-level voice control in ZONOS3-Mini
The most distinctive feature is control over how individual words are delivered, not just how they sound.
Identity controls cover age, gender, accent and voice texture through speaker embedding vectors and accent task vectors. Delivery controls cover pace, emotion, projection and liveliness through activation steering. Learned alignment heads let that steering apply to specific words during streaming, which means a developer can mark one word in a sentence to be delivered emphatically, quietly or quickly while the rest of the sentence flows normally.
A separate edit model, based on a VoiceCraft recipe, can modify sentences after generation, which turns a rendered clip into something closer to an editable document. For anyone who has generated a long speech clip and heard one wrong emphasis near the end, post-generation editing is not a gimmick, it is the difference between usable and regenerating.
These controls also raise the stakes on misuse, because fine-grained delivery control is exactly what makes cloned speech convincing. The model card and the Apache 2.0 licence leave the ethical handling to implementers, and anyone shipping voice cloning owes their users consent mechanisms that the weights alone do not provide. That responsibility has been the industry's sticking point with open TTS for years, and it does not disappear because the model got smaller.
What ZONOS3-Mini makes possible at 10 minutes per generation
The 10-minute generation ceiling and the streaming iSTFT head combine into use cases that clipped TTS could not serve.
Audiobook and narration pipelines can render long passages in a single pass, with word-level steering applied to emphasis as the text streams. Accessibility tools can read documents aloud with consistent voice and controllable pace. Interactive applications can stream speech as it generates, so a response begins playing before the full clip exists, which is what the 80 millisecond time to first audio is really for. And because the alignment heads steer specific words during streaming, the steering can be driven programmatically, say from markup in the input text, rather than hand-tuned clip by clip. Those are the workflows where ZONOS3-Mini's size and licence matter more than its leaderboard position.
None of this requires the 262K context that language models advertise, because speech consumes the context differently. The 32,768 positions are what bound the 10-minute generation, and for most voice work that ceiling is generous.
Four ways to deploy ZONOS3-Mini
The GitHub repository lists four official paths, and their breadth is the practical headline of the release.
The browser path is the boldest. A WebGPU playground runs ZONOS3-Mini fully on the visitor's GPU with no install and no server, and Zyphra reports under 570 milliseconds time to first audio on a Pixel 9 running it in the browser. The NVIDIA server path uses SGLang-Omni to serve an OpenAI-compatible speech endpoint with streaming, which means existing application code can adopt it with minimal changes. The native path includes an MLX server for Apple Silicon, where the model reached under 150 milliseconds to first audio on an M3 MacBook Air, and a C++/ggml server and CLI covering CPU, CUDA, Metal and Vulkan. The hosted path is Zyphra Cloud on H200s.
Each path serves a different builder. The browser runtime suits demos, privacy-conscious tools and products that cannot ship a server component. The SGLang path suits production services that already speak the OpenAI audio format. The MLX and ggml builds suit local-first applications on Macs and mixed hardware. Zyphra Cloud is the path for teams that want the model without operating anything, and its existence tells you the company expects some developers to pay rather than self-host.
Latency figures across those platforms use different hardware and are not directly comparable, as the company notes. The spread still tells a story: 80 milliseconds median time to first audio on a single H100 at 8 concurrent requests on the server side, and real-time enough in a phone browser on the edge side. Tiny local speech models have existed before, including KittenTTS Nano, but ZONOS3-Mini is the first of this quality class to run in a browser with cloning and fine control attached.
Why 100 million active parameters matters
The number to watch in this release is not the 800 million total. It is the 100 million active.
Active parameters are what each generated token actually costs in compute. A dense 8 billion parameter model spends all 8 billion on every token, which is why ZONOS2 needed serious hardware. A sparse model like ZONOS3-Mini spends a tenth of that per token, which is why it runs in a browser. The MoE design trades memory for speed, holding a larger model in storage while activating only a small slice per step, and speech generation with its tight latency budgets is exactly the workload where that trade pays.
The pattern is now visible across the local-AI stack. Models are being redesigned around what modest hardware can activate per token, whether through sparsity, through attention architectures that cut context cost, or through compression that shrinks the weights themselves. ZONOS3-Mini applies the same logic to speech, where the problem is harder because the output must arrive continuously and on time.
For developers, the consequence is concrete. Speech interfaces, accessibility tools, real-time translation and voice agents no longer require a GPU server to sound good. A browser page can carry a cloned, steerable voice, and the weights to build it are Apache 2.0.
Questions readers are asking
Is ZONOS3-Mini really open source?
The weights are released under Apache 2.0, which permits commercial use, modification and redistribution. The training data is not open, so open-weight is the accurate description, and it is one of the most permissive licences among the compared models. Fish Audio S2 Pro, for instance, requires a separate licence for commercial use, which makes ZONOS3-Mini the safer default for commercial speech projects.
How well does voice cloning work?
In speaker-embedding mode, the model posts an 83.4 speaker similarity score on Zyphra's clean set, the best of the compared models. In reference-clip mode, the Capability Index drops to 44.7 from 66.2, so the convenience workflow is measurably weaker than the embedding workflow. Plan for the embedding path if fidelity matters.
Can it run without a GPU?
Yes. The ggml build covers CPU alongside CUDA, Metal and Vulkan, which means the model can run on a laptop without a discrete card, with correspondingly lower speed. The browser runtime uses WebGPU and needs a compatible GPU in the visitor's device.
Should the benchmark numbers be trusted?
Treat them as self-reported. Zyphra built the ZTTS2-Eval benchmark, ran the comparisons and published the explorer, and the company itself flags the cross-model latency figures as not directly comparable. The relative claim, that ZONOS3-Mini leads the open field on its own index, is credible. The absolute quality ranking needs independent runs to settle.
The bottom line
ZONOS3-Mini is the strongest signal yet that high-quality speech generation is moving from the datacentre to the device. An Apache 2.0 model with 100 million active parameters, real-time streaming, word-level delivery control and a browser runtime covers capabilities that required cloud APIs a year ago, and the training budget behind it shows the capability no longer requires frontier-scale compute.
The honest caveats are the self-reported benchmarks, the weaker reference-clip cloning, and audio quality that at least one competitor still leads. None of those changes the strategic picture: speech is becoming a local capability, and the weights to build it are free. The next milestone to watch is independent evaluation, because that is what will tell developers whether the Capability Index leader is the real leader or merely the host's favourite.