Ollama adds Phonon, a text-to-speech model you run locally

A product that reads a reply out loud has usually meant a cloud API. The cloud charges by the second. The key has to leave the user's machine. The last part is now optional on a Mac.

Ollama posted the news on 24 October 2026. The models are called Phonon. They come from a vendor called Gradium. The lead English model is Apache 2.0, so a commercial product can use it.

Four smaller 40-million-parameter models cover French, German, Spanish and Portuguese. A fifth build, phonon:40m-multilingual, covers all five. The post says these 40M models carry a CC BY-NC 4.0 licence. That licence rules out commercial use.

What the post says Phonon can do

Phonon returns speech as a 24 kHz mono audio file. The default is a WAV. It can also return raw 16-bit PCM samples. The model works through an OpenAI-compatible endpoint at /v1/audio/speech. Code written against the OpenAI SDK only has to change its base URL.

The post says a reply can start playing after its first sentence. That works through a streaming response. A Phonon model can also read the output of a chat model, so the two run back to back. Voices are picked with a --voice flag. The post does not say how many voices come with each model.

Users need Ollama 0.40.11 or later. The post names Apple Silicon Macs only. Windows and Linux are listed as coming next, with no date.

The speed figure, and whose figure it is

The post says Ollama's run of Phonon is 35 percent faster than the next fastest engine. The number is from a MacBook Pro with M5 Max. It is the median of 10 runs. Each run made 45 seconds of speech.

The post names the next fastest engine. It is Gradium's own runtime, called ptts. So the comparison is Ollama's own number against the vendor's own baseline. Only Gradium could run this comparison.

The post describes the bit-widths behind the number. Gradium's ptts uses 8-bit weights on the CPU, fp16 on the Neural Engine and f32 on the GPU. Ollama runs the model in bf16. The post also says Ollama's word error rate matches a reference Hugging Face run. That is one check that the speed has not cost accuracy.

What this changes for a voice reply

The first thing a team can do is read the English voice back on the user's own Mac. The licence is Apache 2.0. A paid product can bundle it today.

The four 40M language builds are CC BY-NC. A paid product that needs French or German cannot use them yet. A team in that case waits for Gradium to license them differently, or stays on a cloud voice.

A second use is a demo or a developer build that runs offline. The chat model and the speech model sit on one machine. The OpenAI SDK change is one flag on the client. The limit is Apple Silicon only. A Linux server putting this in production has to wait.

Source

Ollama blog post and the Phonon model page on Ollama.