Gradium opened up Phonon, a text-to-speech model small enough to run in a browser tab, and it passed 200,000 downloads in its first two weeks S¹. The model, about 65 MB, generates speech on a phone or laptop with no server and no network connection. That challenges the assumption that quality voice synthesis needs cloud GPUs.

My read: The download number is self-reported by Hugging Face and may include automated traffic, so I would treat 200,000 as a directional signal rather than a precise adoption count. What I find more telling is the licence split: only the English-only 31m variant is Apache 2.0, while the multilingual 40m model is CC BY-NC. If you need French or German in a commercial product, you are negotiating a per-user pricing deal, not grabbing a free weight. The blog title claims Phonon tops the benchmarks, but the available text cuts off before any leaderboard scores appear, so that claim stays unverified.

Gradium built Phonon on the Pocket TTS lineage and the Mimi codec from Kyutai, using a technique called flow matching to generate audio S¹P⁴. The training was numerically unstable at first and needed several engineering fixes to produce consistent quality across all five supported languages: English, French, German, Spanish and Portuguese S¹. According to the model card on Hugging Face, the model outputs 24 kHz speech and streams audio while still generating P⁴.

Two models, two licences

Phonon ships in two sizes, roughly 65 MB and 75 MB S¹. The smaller Phonon 31m speaks English only and carries an Apache 2.0 licence, meaning developers can use it free in commercial products S¹. The larger Phonon 40m covers all five languages but is licensed CC BY-NC for non-commercial work, with commercial pricing based on user count S¹. More than 100 built-in voices ship with the model, along with voice cloning and voice design for custom voices S¹. It also handles text that trips up many TTS systems, including phone numbers, dates and code snippets S¹.

This is a different bet from the cloud-dependent approach we saw when Google DeepMind's Gemini 3.8 TTS cloned a voice from 30 seconds of audio, which relied on server-side processing. Phonon runs on phones, laptops, single-board computers and directly in a browser tab S¹. The launch demos made the point vividly: an Apple Watch with Wi-Fi off talked users through a coffee brew step by step, an iPhone with no signal counted workout reps, and a robot called Lamp from Autonomous wore a sorting hat and spoke in character S¹. A browser demo turned an offline dino runner game into a radio show hosted by a character called Rex S¹.

What developers can check today

The Hugging Face team built Transformers integration so developers could use Phonon from day one S¹. A separate inference library from Gradium, xn-ptts, has 438 stars on GitHub under Apache 2.0 P³. A repo by developer blixt runs Phonon 31M on WebGPU with word timings, created on 1 November with 22 stars P⁵. Hugging Face also reports that an ExecuTorch conversion runs Phonon on a visitor's GPU in the browser, and that Micdrop brings Phonon to Node.js apps, though these integrations may be early-stage S¹.

The wider trend matters here. As Gemini 4 Carbon topped a recent AI index at 59, the industry split between open and gated models keeps widening. Phonon sits firmly on the open side for English, less so for other languages.

A developer building an offline-first fitness app that talks users through exercises without a data connection could download phonon-31m-en from Hugging Face, wire it through Transformers, and test whether the 65 MB model's latency and quality hold up on a mid-range Android phone. The Apache 2.0 licence means no negotiation for English-only use. If the app needs Spanish or German, the next step is contacting Gradium for commercial terms on the 40m model.

The Open TTS Leaderboard, a new Hugging Face project, now tracks quality across more than 9,000 TTS models on the Hub S¹. Whether Phonon holds the top spot there is a claim the available evidence does not confirm.

Sources: S1 — A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech T · P2 — zeliu98/CloserLook3D · P3 — gradium-ai/xn-ptts · P4 — gradium/phonon-31m-en · Hugging Face · P5 — blixt/phonon

Written from 5 sourced items, 5 of them primary.