Ollama v0.40.9 adds Phonon speech generation on Apple Silicon and d3 decision models

Ollama v0.40.9 can now generate speech with Phonon on Apple Silicon, and apps can stream the audio through the OpenAI-compatible /v1/audio/speech endpoint. The release was published on GitHub on October 22 at 01:33 and is marked as a pre-release. The pull request title (#19227) refers to support for GA Gradium Phonon checkpoints.

Speech with Phonon

ollama run can play text, or a chat model's reply, aloud as the audio is generated. The release notes give this example:

ollama run phonon "Hello from Ollama."

Apps that want audio output can stream it through the /v1/audio/speech endpoint, which follows the OpenAI-compatible format. The notes name Apple Silicon as the supported hardware for Phonon.

d3 decision models

The release also adds d3, which brings vLLM Semantic Router's decision models to Ollama. A user can give a model text, structured state, or an image and ask it to classify, answer yes or no, or score options with probabilities. The d3 family comes in 2B, 4B, 9B, and 27B sizes.

The example from the release notes:

ollama run d3 "A customer says their card was charged twice." --ask "Is this a billing issue?"

Flint on Linux and Windows

Flint now runs on Linux and Windows as well as macOS.

Other improvements

• Streaming OpenAI Responses and Anthropic Messages requests now report generation failures instead of ending silently after partial output.
• ollama launch now passes cloud models' context lengths and capabilities to coding tools, which fixes incorrect context and compaction settings.
• Anthropic-compatible web search now uses the model's supported thinking setting in follow-up requests, preventing avoidable failures.

The full changelog is listed as v0.40.8...v0.40.9. Parameter counts are given only for d3. Pricing, context window sizes, and benchmark scores are not listed for either Phonon or d3.

What this means

The streaming fixes matter for anyone running Ollama behind OpenAI Responses or Anthropic Messages clients. Requests that fail mid-generation are now reported, where before they could end silently after partial output. Developers can treat a truncated response as an error rather than a completed one.

The ollama launch change fixes the context length and capability values that coding tools receive for cloud models. That should stop the incorrect context and compaction settings the notes describe.

Phonon adds a speech output path to the same command line and API surface used for text models. Because the /v1/audio/speech endpoint is OpenAI-compatible, apps written against that interface have a defined way to stream audio from Ollama. For now this is limited to Apple Silicon, and the release carries a pre-release label.

d3 adds a different kind of model to the lineup: one built to classify, answer yes or no, or score options with probabilities rather than produce free-form text. The four sizes, from 2B to 27B, let users choose how much capacity they need for those tasks.