Gradium's open-source, on-device text-to-speech suite — models under 50M parameters that run offline on a phone, laptop, Raspberry Pi, or in the browser, with the English 31m model under Apache 2.0.
Last verified · 2026-10-21 · by Moe Ameen
Gradium Phonon is a family of open-source text-to-speech (TTS) models that run entirely on a device — no cloud, no network. Gradium, a Paris voice-AI company founded in 2025 by researchers from the lab Kyutai, open-sourced the suite on October 20, 2026, replacing an earlier closed ~100M-parameter private beta of the same name with a smaller, openly available set. The defining trait is size paired with speed: the smallest model is about 64 MB and, by Gradium's figures, starts speaking in roughly 7 ms on a laptop CPU.
There are three variants. Phonon 31m is a 31-million-parameter English model (~64 MB, 37 built-in voices) released 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. Phonon 40m multilingual is a single 40m model covering all five languages with a large built-in voice set. Crucially, the 40m variants carry a CC BY-NC license: free for non-commercial use, but commercial use requires a license from Gradium. The only fully-commercial-free option is the English 31m model.
Phonon is built to run anywhere. Gradium targets CPUs (Intel, Arm, WebAssembly), GPUs (CUDA, Vulkan, Metal, WebGPU), and NPUs (Apple Core ML, Snapdragon), and demonstrates the models running on phones, a Raspberry Pi 5, and in the browser — Gradium reports faster-than-real-time performance across all of them. The weights are on Hugging Face and load through Transformers; the code at github.com/gradium-ai/xn-ptts ships Rust, Python (with a CLI), Swift, and JavaScript/WebAssembly integrations, an OpenAI-compatible server, and day-one Apple Silicon support via MLX-Audio, plus LM Studio and Ollama integrations. Voice cloning and voice design are offered free through Gradium Studio and the API.
The honest framing: Phonon is a voice model, not a product. Its latency and word-error-rate claims are vendor-reported with no independent benchmarks yet, so test them on your own scripts. And whatever the numbers, it turns text into an audio file and stops — it writes no script, cuts no video, makes no image or carousel, and publishes nowhere. Everything around the voice is yours to build or bring.
Phonon's real edge over a hosted voice API is that it is embeddable and portable: you can drop it into your own pipeline — a laptop CPU, a phone, a Raspberry Pi, even a browser tab via WebAssembly — and voice text locally for nothing. For a creator that unlocks two things a cloud TTS can't: a faceless-video narration layer with zero marginal cost, and, through the 40m line, the same script voiced in five languages for localized versions of a post. But embeddable audio is still only audio. Phonon doesn't decide what to say, doesn't cut the video around the voice, and doesn't get any of it in front of an audience. That is the half [Kompozy](/) owns, and it's the larger half.
Here's how they fit: Kompozy generates the source — the blog article, the narration script, the text posts, the carousels, the newsletter — all held to one voice by the Persona Brief and banned-word filters, so the script you pipe into Phonon is already on-brand. You voice that script locally with Phonon (the English 31m for free commercial use, or a licensed 40m model for a French, German, Spanish, or Portuguese edition), assemble it into a faceless video or a listen-along blog, and hand the rest to Kompozy: it fans the same idea into Clipped and Persona Shorts, brand-exact Carousel Posts, Quote Graphics, Photo Posts, and a blog and newsletter, then schedules and publishes the whole set across the eight social platforms plus blog and email with Autopilot. For talking-head social video Kompozy already renders avatar clips with HeyGen's native TTS, so Phonon stays in its lane — your free, local, embeddable audio track — while Kompozy does generation and distribution.
Phonon is an open-source suite of on-device text-to-speech models from Gradium, a Paris voice-AI company, released October 20, 2026. The models are all under 50 million parameters and run entirely on a device — phone, laptop, single-board computer, or browser. The English Phonon 31m model is Apache 2.0 licensed; the 40m and multilingual variants (English, French, German, Spanish, Portuguese) are CC BY-NC.
The English Phonon 31m model is free under Apache 2.0, including for commercial use. The 40m per-language and multilingual models are CC BY-NC — free for non-commercial use, but commercial use requires a license from Gradium. Local inference and voice creation are free and unlimited across the suite.
Yes. Phonon is designed to run on-device across CPUs, GPUs, and NPUs — including phones and a Raspberry Pi 5 — and it runs 100% client-side in the browser via WebAssembly or WebGPU. Gradium reports faster-than-real-time performance, with first audio in about 7 ms on a laptop CPU (a vendor-measured figure).
The free Apache 2.0 31m model is English only. The 40m line covers English, French, German, Spanish, and Portuguese — either as one model per language or a single multilingual model — but those variants are CC BY-NC, so commercial use of their audio requires a license from Gradium.
Use Kompozy to generate the content — the blog, the script, the posts — under your brand voice and to publish it across nine platforms, then run the written script through Phonon locally to voice a free narration track for a faceless video or a listen-along channel. Kompozy owns generation and distribution; Phonon owns the free, local, on-device audio layer.