What happened
Ollama’s GitHub release notes list v0.40.11 as released on October 24, 2026. The headline addition is speech generation on macOS. According to the notes, ollama run can play supplied text or a chat model’s reply aloud while audio is being generated. The documented example invokes Phonon with ollama run phonon "Hello from Ollama." This describes playback as part of the command-line workflow rather than requiring a user to first generate and manage a completed audio file.
For application developers, the release notes say applications can stream audio through an OpenAI-compatible /v1/audio/speech endpoint. That gives software built around an OpenAI-style audio API a defined integration surface for speech output through Ollama. The announcement says the audio can be streamed, but it does not include a request example, authentication details, endpoint parameters, supported output formats, or a list of client libraries.
The release also introduces d3 decision models from vLLM Semantic Router to Ollama. The notes say a d3 model can receive text, structured state, or an image, then be asked to classify something, give a yes-or-no answer, or score options with probabilities. Ollama lists four sizes in the family: 2B, 4B, 9B, and 27B. Its example presents a statement that a customer was charged twice and asks whether that is a billing issue. The example frames d3 as a model family for bounded decisions rather than open-ended chat generation.
This is one release with two distinct capabilities. Speech output is explicitly described as a macOS feature, while d3 broadens the model choices available through Ollama. The release page records 39 commits to main since v0.40.10 and links the full changelog for the two versions. It does not announce a new price, a separate access tier, or usage limits for either capability. It likewise does not say whether the speech and decision-model additions share particular runtime, hardware, or deployment requirements.
What the archive adds
New Runtime’s earlier coverage described Phonon 0.6.0 adding 13 languages, restoring Metal support on macOS 27, and enabling Core ML export: Release v0.6.0: gradium-ai/xn-ptts, dated November 1, 2026. An earlier note covered a subsequent Phonon release with smaller models, added languages, and custom-voice support: A Closer Look at Phonon TTS, dated November 5, 2026. This is a new Ollama integration and access point for speech generation, not a new Phonon model announcement.
Trend signal
This supports generative media becomes application infrastructure: speech is exposed through an application-callable endpoint rather than only as a standalone generation experience. The macOS command-line playback path and streaming endpoint make speech a component a developer can wire into a product workflow. The release establishes the interface, but does not demonstrate that it is inexpensive, low-latency, scalable, or suitable for production workloads.
Why now
Ollama places speech generation beside its existing local model-running workflow and an OpenAI-compatible API surface. That combination makes the change immediately relevant to developers who already use Ollama as a local runtime and want to test spoken output without redesigning an audio integration. The release notes do not explain a broader product timetable or identify a particular customer demand that prompted the addition.
Why it matters
For teams prototyping voice-enabled software, the practical change is not merely that a speech model can run locally. Ollama now describes a path to play generated audio directly in a macOS command-line workflow and a path for an application to receive streamed speech through a familiar API shape. A team with an existing OpenAI-compatible client may therefore be able to test a local speech-output path with less integration work than a wholly separate service would require.
That does not make this a demonstrated replacement for a hosted text-to-speech provider. Builders will need to assess whether the output, responsiveness, supported hardware, and operational behavior fit their own product. The d3 addition has a different potential use: it offers sizes from 2B through 27B for routing, classification, and probability-scoring decisions where a structured answer is more useful than a free-form response. Both additions reinforce the value of treating local model infrastructure as a set of callable application components, while keeping evaluation responsibility with the team deploying them.
Limits and uncertainty
Ollama’s announcement does not give hardware requirements, supported macOS versions, measured generation speed, audio quality, or language and voice coverage for speech generation. It does not specify endpoint parameters, formats, concurrency limits, or whether the API is intended for production workloads. The release identifies Phonon in its command example but provides no model size, licensing detail, or performance results. It also does not report d3 accuracy, evaluation methodology, task boundaries, or comparative results. These are release-note claims and have not been independently tested here.
Skill to build now
Build a small macOS proof of concept that sends one fixed sentence and one real chat response through /v1/audio/speech, then record time to first audible output and the behavior when requests overlap. Run the same prompts through ollama run phonon so the command-line path and application path are tested separately. For d3, choose one existing routing decision with a known correct label, such as support-ticket categorization, and compare the 2B and 9B variants before assigning the model an automated action.
Reader takeaway
Ollama v0.40.11 makes speech generation callable from macOS workflows and exposes streaming through an OpenAI-compatible endpoint. It is a concrete local integration path worth testing, not evidence yet that local speech can replace a hosted service. The same release also adds d3 models for classification, binary decisions, and probability scoring, without publishing evaluation results.