The Pulse
Ollama Adds Gradium’s Stable Phonon Checkpoints in v0.40.9
Ollama v0.40.9 adds support for Gradium’s GA-format Phonon speech checkpoints on Apple Silicon, replacing the earlier prerelease checkpoint path. The release also enables speech generation from the CLI and through an OpenAI-compatible API e
AI.info Team · Written with AI · fact-checked ·
Ollama’s new support for Gradium’s stable Phonon checkpoints comes with a compatibility boundary: the release replaces the earlier preview checkpoint path, rather than simply extending it. Version 0.40.9, published October 22, recognizes Gradium’s production-format speech models and makes that version the minimum for Ollama’s packaged Phonon models.
From preview files to Gradium’s GA format
The change is documented in Ollama’s implementation commit, which describes replacing the prerelease checkpoint path with Gradium’s GA format. The update adds handling for the checkpoint’s default voice data and speech-conditioning inputs, alongside tests for importing the new format and rejecting the old checkpoint. Ollama’s release notes confirm the user-facing result: Phonon speech generation on Apple Silicon.
That format change matters to developers bringing their own checkpoints into Ollama. The implementation identifies Gradium’s model architecture as GradiumPhononTTSForConditionalGeneration and imports the accompanying default-voice.safetensors file. The update also raises the documented minimum Ollama version from 0.40.8 to 0.40.9, so the new checkpoint path depends on a runtime upgrade.
The importer requires the default-voice file and checks that it contains the expected speaker_wavs tensor before mapping it to the model’s default voice. The commit adds tests for the GA architecture, voice-file validation and rejection of the prerelease architecture. It also keeps Phonon’s weights unquantized, another constraint for developers preparing checkpoints to import.
Speech streams through the CLI and API
Ollama says ollama run can play supplied text—or a chat model’s reply—aloud as the audio is generated. Apps can also stream generated audio through the OpenAI-compatible /v1/audio/speech endpoint. The change gives local applications a speech-output route within Ollama, rather than limiting Phonon use to a separate runtime.
Users can start with ollama run phonon or connect a chat response to the speech model. The Ollama model listing describes Phonon as Gradium’s text-to-speech family and says it produces 24 kHz mono audio. The packaged models require Apple Silicon, a limitation that narrows the release’s immediate reach to Apple’s Arm-based systems.
Model sizes bring different language and license terms
Ollama’s default Phonon model is an English checkpoint with 31 million parameters and a voice called default. The library also lists 40-million-parameter models for English, French, German, Spanish and Portuguese, plus a multilingual model covering those five languages. Ollama reports loaded memory of 121 MB for the 31M model and 141 MB for the listed 40M variants on an Apple M5 Max.
Licenses differ by checkpoint: the 31M English model uses Apache 2.0, while the 40M language-specific and multilingual models use CC BY-NC 4.0, which restricts commercial use. Developers therefore need to check the model’s license as well as its language and hardware requirements before integrating it. The version boundary is explicit: Ollama’s Phonon packages need Apple Silicon and Ollama 0.40.9 or later.