Ollama adds Phonon speech: small models, important limits

Ollama has added streaming text-to-speech, starting with Gradium’s Phonon family. For local-AI builders, the useful change is a familiar runtime gaining a speech output interface. Two details matter before adopting it: the supported platform and the exact checkpoint’s license.

Event: Ollama’s announcement is dated October 24, 2026. This article was published and sources reviewed October 25, 2026.

What Ollama added

The October 24 announcement introduces speech generation on Apple Silicon Macs, with Windows and Linux support described as forthcoming. It recommends Ollama 0.40.11 or later. The broader device coverage of Gradium’s own runtime should not be mistaken for support in these particular Ollama packages.

Ollama attributes its Apple Silicon acceleration to MLX and publishes an M5 Max comparison. We have not reproduced that benchmark. Audio throughput is not a text-model tokens-per-second result, and it does not establish performance on our RTX 3060 system.

Small weights do not mean the whole app fits in that space

The Phonon 31M English model card lists 64.5 MB for its quantized weights and notes that the parameter count excludes the audio tokenizer. That is a weights-file measurement, not a complete application, download or memory budget. The card describes streaming 24 kHz mono speech.

The same card distinguishes the English 31M checkpoint’s Apache 2.0 license from the CC BY-NC 4.0 terms for the 40M checkpoints, including the multilingual variant. “The Phonon family” is not one blanket permission for commercial use.

An audio endpoint, not another chat model

The speech documentation describes streaming through /v1/audio/speech, with WAV and raw PCM output. It sets a 4,096-character API input limit; the CLI can split longer passages. Speech models reject chat, completion and embedding requests. That distinction matters for launchers that currently send every selected model through a text-chat route.

Voice options should be checked against the installed package. The documentation says unsupported options return errors, so an interface should expose capabilities rather than assume every speech backend accepts the same controls.

Our take: test the complete reply

For an assistant, a useful acceptance test runs from a user’s question to the first audible reply, then checks cancellation, pronunciation and resource use while other work continues. Speech generation alone does not provide listening, conversation reasoning or interruption handling. A smaller voice component could leave more room for those parts; that is a design opportunity, not a measured result here.

AiBrainWorX test status: source review only. No runtime or model weights were installed or downloaded. This article does not announce built-in Phonon support in our launcher. Explore our model library for separately documented options.