Ollama Phonon on a Mac: 104MB TTS, 121MB Loaded on M5 Max

Local speech on a Mac does not need a multi-gigabyte model. Phonon is Gradium's text-to-speech family, packaged in Ollama as 14 tags that download between 104MB and 148MB. Ollama publishes the loaded memory on an Apple M5 Max right on the model page: 121 MB for the 31M English checkpoint and 141 MB for every 40M checkpoint. The packages need an Apple Silicon Mac and Ollama 0.40.9 or later, and the smallest tag is speaking after one command.

The catch is not hardware, it is licensing. Only the 31M English checkpoint is Apache 2.0; every 40M checkpoint, all the non-English languages included, is CC BY-NC 4.0 and therefore noncommercial only.

Does Phonon run on your Mac?

Yes, on any Apple Silicon Mac. The largest tag in the family is 148MB on disk. The heaviest checkpoint loads 141 MB of memory on an Apple M5 Max, which is a rounding error next to the unified memory of the smallest Mac Apple sells today. There is no VRAM tier to consult and no model size to divide by a quantization factor, because a text-to-speech checkpoint of this size runs in the same unified memory pool as any other Ollama model.

The four things worth checking before you pull:

• Chip: Apple Silicon only. The library page states the packages require Apple Silicon, and the launch post says speech models are available for macOS, with Windows and Linux support coming soon.
• Ollama version: the library page names Ollama 0.40.9 or later. The October 24 blog post instead tells you to download Ollama 0.40.11 or later and states that its published measurements used 0.40.11. Install 0.40.11 or later and the two agree.
• Disk: 104MB for the English 31M tags, up to 148MB for the multilingual 40M tag.
• Licence: Apache 2.0 for English 31M, CC BY-NC 4.0 for the rest. That one decides whether you can ship it, so it gets its own section below.

If you have not installed Ollama yet, the Mac install walkthrough covers the brew and direct-download paths, and the runtime comparison explains how Ollama accelerates on Apple Silicon.

The 14 tags, and the two columns people confuse

A tag is the exact Ollama name you pull, such as phonon:fr, and several tags can resolve to the same checkpoint. Download size is how much Ollama downloads for that tag. Loaded memory is what the checkpoint occupies in unified memory while it runs.

The model page shows one table with a download size per tag. Further down, the README carries a different table with the loaded memory per checkpoint on an Apple M5 Max. They are not the same number and they answer different questions. Nothing in the family downloads 121 MB: 121 MB is what the 31M English checkpoint occupies once loaded on that machine, and 141 MB is the same figure for every 40M checkpoint.

Two details in that table are worth reading twice. First, phonon:latest is listed at 113MB while its own aliases phonon:en and phonon:31m-en are listed at 104MB, and the README names all three as the 31M English checkpoint. Second, the short language aliases and their numbered counterparts do not carry the same listed size either: phonon:fr is 123MB and phonon:40m-fr is 126MB, for the same French checkpoint. Read the size from the exact tag you pull rather than from the family name.

A checkpoint is one set of trained weights, and the 14 tags in that table resolve to 7 of them: the 31M English checkpoint under three tag names, and six 40M checkpoints under the other eleven. Counting the tags overstates the family. Counting the checkpoints gives the number that decides both your licence and your language coverage.

Every tag in the list carries the same 8K context window and takes text input. Phonon produces 24 kHz mono audio, and the 31M English checkpoint ships a single voice, named default.

Setup on a Mac

Install Ollama 0.40.11 or later, then pull the tag you want. The English 31M tag is the smallest and the only one you can use commercially.

ollama pull phonon:31m-en
ollama run phonon:31m-en "Hello from Ollama."

Voices come from Gradium's voice catalog, and --voice picks one. The launch post notes that ollama show phonon lists the voices a checkpoint carries.

ollama run phonon:31m-en --voice Zoey "So I went through the whole thread. Short version? Everyone agrees, and nobody wants another meeting."

Redirecting the output writes a WAV file. A chat model's reply can also be piped straight into the speech model, so playback starts after the first sentence.

ollama run phonon:31m-en "Save this to a file." > speech.wav
ollama run gemma4 --hidethinking "Tell me a fun fact about owls in two sentences." | ollama run phonon:31m-en

For code that already talks to the OpenAI SDK, the endpoint is OpenAI-compatible at /v1/audio/speech and streams as it generates.

curl http://localhost:11434/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model":"phonon:31m-en","voice":"Zoey","input":"Hello from Ollama."}' \
  --output speech.wav

What the 35% faster figure covers, and what it does not

The only speed number Ollama published is its own, and it is narrow. On a MacBook Pro with an M5 Max, Ollama was 35% faster than the next fastest engine, Gradium's own ptts runtime, measured generating 45 seconds of speech as a median of 10 runs. ptts is Gradium's own inference runtime, and it is the comparison point for that figure. Accuracy is stated in the same breath: the word error rate matches the reference implementation in Hugging Face Transformers on 200 sentences from the Seed-TTS English test set, transcribed with Whisper large-v3. For context on the two paths, Ollama runs the model in bf16, while ptts uses 8-bit weights on the CPU, fp16 on the Neural Engine and f32 on the GPU.

Three limits keep that honest. It is one machine and one engine pair, not a throughput number you can transfer to a MacBook Air. There is no published per-tag table, so you cannot compare 31M against 40M on your own chip from the sources. And ModelFit has run no test on Phonon: every figure here is opened in a publisher's page, and no measurement of our own is claimed.

The licence split decides whether you can ship it

• The 31M English checkpoint is Apache 2.0. Use it commercially, including in a paid app.
• Every 40M checkpoint is CC BY-NC 4.0, which limits it to noncommercial use. That covers French, German, Spanish, Portuguese, English and the multilingual checkpoint.

Apache 2.0 is a permissive licence that allows commercial use. CC BY-NC 4.0 is the Creative Commons variant whose NC clause restricts the same work to noncommercial use.

So a commercial product that needs non-English speech is outside what these Ollama packages grant, whatever the machine. English-only projects pay for nothing: the Apache 2.0 tag is also the smallest download of the family at 104MB.

Which tag to pull, by job:

For example, an app that ships English narration commercially takes phonon:31m-en: the licence permits it and 104MB is the smallest download in the family. A hobby project that wants the same paragraph in five languages pulls phonon:40m-multilingual for 148MB, inside the noncommercial terms.

Pocket TTS voice support in llama.cpp, and why it is not a new model

The second half of the story is on the llama.cpp side. Pull request 31429, "Pocket tts - support for voices without voice cloning", merged on November 2, 2026, and the first nightly to list it is b12061, published the same day with a macos-arm64 binary.

Pocket TTS is the Kyutai text-to-speech model that Phonon builds on, and the Mimi codec comes with it; the Pocket TTS weights date from January 2026. What shipped in b12061 is runtime support for the predefined-voice pack, where each voice is stored as backbone state after the voice prompt, converted into the audio projector and exposed under voices in GET /props. The converter warns when a shipped voice state drifts above 5% from the weights it is paired with.

Voice cloning is the ability to reproduce a speaker from a reference clip. Two things that PR does not give you. It is the without-voice-cloning path, so predefined voices only, no cloning: a speaker reference is rejected when the model ships no voice encoder. And it is a nightly build number, not a release, so treat b12061 as a floor rather than a stable version.

Mac compatibility by memory tier

Phonon changes no machine decision, which is the useful finding here. Any Apple Silicon Mac runs it, and the size difference between tiers is irrelevant.

The practical rule for the rest of the machine is unchanged: a speech model this small coexists with a chat model in memory. Pairing Phonon with a quantized 8B to 14B model on a 16GB Mac costs you almost nothing. For a worked example of a small Ollama model on Apple Silicon, including the setup and the RAM budget, see LFM2.5 on Mac. For what does and does not fit at each memory size on the language-model side, the RAM tier guide is the reference.

FAQ

Is Phonon a new model?

Phonon is a new family from Gradium, announced on October 24, 2026, but its technical base is not new. The model card states that Phonon builds on Pocket TTS and the Mimi codec from Kyutai, and the Pocket TTS weights date from January 2026. What changed in the last week is packaging and runtime support, not the underlying research.

Does Phonon run on an Intel Mac, Windows or Linux?

No. The library page states the packages require Apple Silicon, and the launch post says speech models are available for macOS with Windows and Linux support coming soon. On an Intel Mac the documented Ollama path does not apply.

Can I clone a voice with Phonon in Ollama?

Not with these packages. The 31M English checkpoint ships one voice named default, and the llama.cpp work merged on November 2 is explicitly the without-voice-cloning path, where predefined voices are supported but a speaker reference is rejected when the model has no voice encoder. Voice cloning and voice design sit in Gradium's own studio, not in the Ollama tags.

Can I use the 40M models in a commercial product?

No. The 40M English, French, German, Spanish, Portuguese and multilingual checkpoints are CC BY-NC 4.0, which limits them to noncommercial use. Only the 31M English checkpoint, at Apache 2.0, is clear for commercial work.

How much disk and memory does Phonon need?

Disk is 104MB to 148MB depending on the tag. Loaded memory is 121 MB for the 31M English checkpoint and 141 MB for every 40M checkpoint on an Apple M5 Max. Those two figures come from two different tables on the same model page, the disk number per tag from one and the loaded memory per checkpoint from the other.

Sources

• Ollama, Phonon model page, read November 4, 2026: the 14 tags, their download sizes, the loaded memory on Apple M5 Max (121 MB and 141 MB), the Apache 2.0 and CC BY-NC 4.0 split and Ollama 0.40.9.
• Ollama, Speech models in Ollama, October 24, 2026: the 35% figure against ptts, the 45 seconds of speech as a median of 10 runs on a MacBook Pro with M5 Max, and the 200 Seed-TTS word error rate check.
• Gradium, Phonon 31M English model card: the 31M parameter count, the Apache 2.0 licence and the 24 kHz mono output, and the statement that Phonon builds on Pocket TTS and the Mimi codec.
• Kyutai, Pocket TTS: the January 2026 date of the Pocket TTS weights.
• kyutai/pocket-tts-without-voice-cloning: the predefined-voice pack the llama.cpp work targets.
• llama.cpp pull request 31429, "Pocket tts - support for voices without voice cloning", merged November 2, 2026: the predefined-voice support, the voice listing in GET /props and the rejection of a speaker reference without a voice encoder.
• llama.cpp release b12061, published November 2, 2026: the build number, the macos-arm64 binary, and the release notes read through the GitHub API record.
• Ollama registry probes for all 14 library/phonon tags, every one returning HTTP 200, checked November 4, 2026.

Every number here is a figure opened in a publisher's page on November 4, 2026. ModelFit ran no benchmark on Phonon and claims no Apple Silicon throughput figure of its own. How we work: about ModelFit.