The Paris voice-AI company says its smallest Phonon model starts speaking in about 7 ms on a laptop CPU and runs entirely on a phone, laptop, single-board computer, or browser — with the 31M English model released under Apache 2.0 and larger multilingual variants under a non-commercial license.
2026-10-21 · by Moe Ameen
Gradium, a Paris-based voice-AI company founded in 2025 by researchers from the lab Kyutai, announced on October 20, 2026 that it is open-sourcing Phonon — a family of text-to-speech (TTS) models small enough to run entirely on a device, with no cloud and no network connection. It is a notable shift: Gradium has to date sold real-time, server-side voice APIs (streaming speech recognition, synthesis, voice cloning, translation) and earlier described Phonon as a roughly 100M-parameter private beta. The open release is a different, smaller set of models.
The suite has three variants, all small. Phonon 31m is a 31-million-parameter English model, about 64 MB at 8-bit quantization, shipping with 37 built-in voices under the permissive Apache 2.0 license. Phonon 40m is a 40-million-parameter, roughly 75 MB model offered one-per-language across English, French, German, Spanish, and Portuguese, plus a single 40m multilingual model covering all five. The 40m variants carry a CC BY-NC license, meaning commercial use requires a separate license from Gradium — so the fully commercial-free path is the English 31m model. The architecture is a compact transformer (6 layers for 31m, 9 for 40m) that outputs 24 kHz audio.
Gradium's headline claim is latency: it reports the 31m model begins speaking in about 7 ms on a MacBook Air CPU and runs many times faster than real time, with similar sub-10 ms figures on a Snapdragon 8 Elite phone and faster-than-real-time performance on a Raspberry Pi 5. It also reports lower word-error rates than other small open models (Kokoro among them) on a set of deliberately hard sentences — phone numbers, emails, postal codes. These are vendor-reported numbers without independent third-party benchmarks yet, so treat them as the company's claims and test on your own scripts.
The weights are on Hugging Face and load through Transformers; the code lives at github.com/gradium-ai/xn-ptts with Rust, Python (CLI), Swift, and JavaScript/WebAssembly integrations, an OpenAI-compatible server, and day-one Apple Silicon support via MLX-Audio. The models target CPUs (Intel, Arm, WebAssembly), GPUs (CUDA, Vulkan, Metal, WebGPU), and NPUs (Apple Core ML, Snapdragon), and Gradium demonstrates them running in the browser. Local inference and voice creation are free and unlimited; custom models and commercial use of the non-commercial variants route through Gradium sales.
The move for a working creator this week is narrow and real: the English Phonon 31m is Apache 2.0, so you finally have a commercially-usable, free, offline voice you can run on a laptop CPU as many times as you like — a genuine narration layer for a faceless channel. But a WAV file is not a channel. Phonon reads text; it writes none, cuts no video, and posts nothing. [Kompozy](/) is the engine that supplies the words Phonon reads and ships everything around them. Kompozy writes the [blog article](/glossary/content-repurposing), the narration script, the text posts, and the email newsletter under a [Persona Brief](/glossary/persona-brief) and banned-word filters, so what you feed into Phonon is already on-brand and audience-fit before the first syllable.
The concrete loop: draft one idea, let Kompozy write the long-form article and a tight script in your voice, run that script through Phonon locally at zero marginal cost for a clean voiceover, then let Kompozy fan the same source into the pieces a voice model can't make — [Clipped and Persona Shorts](/glossary/persona-shorts) for the feeds, brand-exact carousels, quote graphics, photo posts, plus the blog and newsletter — and schedule and publish the set across the eight social platforms plus blog and email from one queue with [Autopilot](/glossary/autopilot). For talking-head social video Kompozy already generates avatar clips with HeyGen's native TTS, so the two rarely overlap: Phonon owns your standalone, free, local audio track, and Kompozy owns generation and cross-platform distribution.
Phonon is a suite of open-source, on-device text-to-speech models from Gradium, a Paris voice-AI company, announced October 20, 2026. The models run entirely on a device (phone, laptop, single-board computer, or browser) with no cloud. The smallest is Phonon 31m, a 31-million-parameter English model under the Apache 2.0 license; larger 40m and multilingual variants cover English, French, German, Spanish, and Portuguese.
Partly. The English Phonon 31m model is released under Apache 2.0, so you can run it and use its audio commercially for free. The 40m per-language and 40m multilingual models carry a CC BY-NC (non-commercial) license, so using their output in monetized content requires a separate license from Gradium. Local inference and voice creation are free and unlimited.
Both are tiny open TTS models that run locally on a CPU, and Gradium positions Phonon directly against Kokoro. Gradium reports faster first-audio latency and lower word-error rates on hard text for Phonon, but those are vendor-measured figures without independent benchmarks. Kokoro is Apache 2.0 across the board, while only Phonon's 31m model is; benchmark both on your own scripts before committing.
No. Phonon turns text into audio and nothing more — it writes no script, makes no video or images, and posts to no platform. To go from a Phonon voice track to published content you need a content engine like Kompozy, which generates the copy, video, carousels, blog, and newsletter and publishes them across nine platforms.