In short
- Phonon is Gradium's TTS family, on the Ollama library as of 24 October 2026 — text in, 24 kHz mono audio out, generated as it goes.
- It's tiny: 104–148MB downloads, 121–141MB loaded memory on an Apple M5 Max — any Mac that runs Ollama can run it.
- Mac-only for now: requires Apple Silicon and Ollama 0.40.9 or later; no other engines documented as of release day.
- Licence split matters: 31M English is Apache 2.0; every 40M checkpoint is CC BY-NC 4.0 (non-commercial).
- No published quality numbers yet — judge it by ear against Piper or Coqui before committing.
If you searched "phonon ollama", here's the short answer: phonon is Gradium's text-to-speech family, and it's on the Ollama library as of today, 24 October 2026. You give it text, it gives you 24 kHz mono audio — spoken as it generates, not after. It's not a chat model, and at 104–148MB per download it's smaller than almost anything else you'd install through Ollama.
What phonon is
Phonon is a TTS family from Gradium. It turns written text into spoken audio as it generates, which the Ollama page says makes it useful for three things: reading assistant replies aloud, narrating passages, and giving a voice to local apps.
That's the whole job. There's no chat, no tools, no thinking mode — text in, audio out. If you've built a local assistant and want it to actually speak its answers rather than print them, this is the kind of model you'd put on the output side.
It also fits a pattern we've been covering all month: the Ollama library taking on models that don't generate text at all. Decision models like kev and pplx-decider, embedders like EmbeddingGemma 2 — and now a speech model. Ollama is quietly becoming a place you install small, single-purpose models with one command, not just chat LLMs.
Sizes, languages and memory
There are two checkpoint sizes:
- 31M — English only. This is the default.
- 40M — English, French, German, Spanish and Portuguese, plus a multilingual checkpoint covering all five.
The downloads are tiny by any standard:
| Ollama tag | Language | Download | Loaded memory (Apple M5 Max) |
|---|---|---|---|
phonon (phonon:31m-en, phonon:en) | English, 31M | 104MB | 121MB |
phonon:40m-en | English, 40M | 132MB | 141MB |
phonon:40m-fr (phonon:fr) | French, 40M | 126MB | 141MB |
phonon:40m-de (phonon:de) | German, 40M | 126MB | 141MB |
phonon:40m-es (phonon:es) | Spanish, 40M | 129MB | 141MB |
phonon:40m-pt (phonon:pt) | Portuguese, 40M | 125MB | 141MB |
phonon:40m-multilingual (phonon:multilingual) | All five, 40M | 148MB | 141MB |
(Loaded-memory figures are the readme's own measurements on an Apple M5 Max.)
Our usual rule of thumb — budget roughly the download size in RAM or VRAM plus context headroom — holds here, except the numbers are so small the rule barely matters. All variants have an 8K context window and take text input only. To put it in perspective: an 8GB machine, about the minimum for a comfortable Ollama setup, has roughly 60× the memory the 31M model loads into.
What hardware you need
Memory is a non-issue. Any machine that can run Ollama has the RAM phonon needs many times over.
The catch is the platform: these Ollama packages require Apple Silicon and Ollama 0.40.9 or later. So despite the tiny footprint, it's Mac-only for now. There's no gaming-GPU route, no Linux box, no Windows — at least as of today. That's a situation we've seen before this month: kev and strands-decider both shipped as Mac-only Ollama releases.
We've found no LM Studio, llama.cpp or vLLM instructions on the phonon page either. The readme links to Hugging Face checkpoints for each model, but the Ollama page is the only documentation we've reviewed — we haven't tested those checkpoints outside Ollama. If you're not on a Mac, there's no documented way to run it locally today.
If you are on a Mac, hardware class is irrelevant in the good sense: a base MacBook Air handles a 31–40M model without breaking a sweat. There's no reason a model this small would need a discrete GPU.
How to run it
The default pull is Phonon 31M English, with one voice, named default:
ollama run phonon "Good morning. Here is a short update before your first meeting."
That's the example from the readme, and it's genuinely the whole setup — one command, audio out.
For other languages, pick a tag:
ollama run phonon:40m-de # German
ollama run phonon:40m-fr # French
ollama run phonon:40m-es # Spanish
ollama run phonon:40m-pt # Portuguese
ollama run phonon:40m-multilingual # all five
Short aliases exist too — phonon:de, phonon:fr and so on — and they point at the same 40M checkpoints.
Two practical notes from the readme:
- Feed it text meant to be spoken. Formatting marks like Markdown headings may be read aloud. If you're narrating assistant replies, strip the formatting first — otherwise your voice assistant may cheerfully read out your
##symbols. - For saving audio or using the speech API, the readme points to Ollama's speech guide — that's what you'll want for any real app rather than one-off CLI runs.
If you're new to Ollama itself, our engine guide covers setup, and the LM Studio vs Ollama piece explains why Ollama is the natural fit for one-command installs like this.
Licences: check before you build
This is the part worth reading twice, because the licence split is unusual within a single family:
- The 31M English checkpoint is Apache 2.0. Use it commercially however you like.
- Every 40M checkpoint — English, French, German, Spanish, Portuguese and multilingual — is CC BY-NC 4.0. Non-commercial only.
The practical consequence: if you're building anything that makes money, you're on the single-voice 31M English model. The other languages and the multilingual checkpoint are off-limits for paid products.
It's a stranger arrangement than most families we've covered recently — bonsai-2 shipped under Apache 2.0 across all four sizes. Phonon effectively gives away the small English model and gates everything else behind non-commercial terms — fine for hobby projects, a trap for anyone who doesn't read to the bottom of the readme.
What it's good for, and its limits
Where it fits:
- Reading local assistant replies aloud. Pair it with a small chat model — qwen3:1.7b or smollm2:360m — and the whole voice-assistant pipeline still fits in a few GB. That's the headline use case, and the "as it generates" design is aimed at it: speech that starts flowing while the text is still being processed suits live assistant replies better than batch narration.
- Narrating documents and passages. The 8K context window caps very long single passages, but for typical articles and replies it's ample.
- Prototyping voice features in local apps, via Ollama's speech API.
The limits:
- One voice on the default model, named
default. The page doesn't list voice options for the 40M models, so don't expect a voice picker. - 24 kHz mono is fine for speech but not studio quality.
- Mac-only via Ollama, as above.
- No capabilities are listed on the Ollama page and no benchmarks are given. We haven't seen independent quality comparisons against other local TTS options, so treat this as a first look, not a verdict.
[[cta:host]] — a 31M TTS model sips memory: a Mac running phonon alongside a chat model has capacity to spare, and the Inpherio connector can serve requests when the machine is idle.
How it fits the local TTS landscape
Local TTS is already well served. Piper is the established go-to for tiny, fast CPU speech; Coqui TTS and Kokoro-style models offer more voices and richer output. Ollama itself now has a speech API, which phonon plugs straight into.
Phonon's differentiators, as presented: it's on the Ollama library with one command, it generates audio as it goes, and it's genuinely tiny (104–148MB). If your stack is already Ollama-shaped, that's a real convenience — one engine, one API, chat and speech together.
What we can't say yet is how it sounds next to Piper or Coqui. No quality numbers are published, so we can't rank it on naturalness — try it and judge by ear. For a first release with no benchmarks, that's the honest position.
If you want TTS without a Mac, the marketplace has hosted options callable per hour of audio — chatterbox-turbo at £0.41/hour or Qwen3-TTS at £0.41/hour, both through the same API.
What we'd pick: on a Mac, ollama run phonon costs you 121MB of memory and a single command — try it before reaching for anything heavier. For commercial work, stick to the 31M English checkpoint and mind the licence. Off Apple Silicon, Piper remains the sensible default until phonon lands elsewhere.
Questions people ask
How do I run phonon on Ollama?
ollama run phonon for the default 31M English model, or a tag like ollama run phonon:40m-de for German. You need Ollama 0.40.9 or later on Apple Silicon.
How much memory does phonon need?
Per the readme, 121MB loaded for the 31M English model and 141MB for the 40M variants on an Apple M5 Max. Downloads are 104–148MB. Any Mac that runs Ollama can run it.
Can I use phonon commercially?
Only the 31M English checkpoint, which is Apache 2.0. The 40M English, French, German, Spanish, Portuguese and multilingual checkpoints are CC BY-NC 4.0 — non-commercial use only.
Does phonon work on Windows, Linux or NVIDIA GPUs?
Not via Ollama as of 24 October 2026 — the packages require Apple Silicon. We've found no other engine instructions on the model's page.
What languages does phonon support?
English (31M and 40M), plus French, German, Spanish and Portuguese at 40M, and a 40M multilingual checkpoint covering all five.
Models in this post — live on Inpherio
Have a GPU, a Mac or a spare PC?
Let it earn while you sleep.
Run open models for people who need them. You set the price and keep everything but the 15% fee.
Become a hostBuilding with AI?
Every open model, one API.
OpenAI-compatible. Priced by the people who run the hardware.
Start free See the marketplaceRead “Phonon on Ollama: Gradium's tiny text-to-speech family, first look” at https://inpherio.co.uk/blog/phonon-on-ollama-gradium-s-tiny-text-to-speech-family-first-look