Ollama can now generate speech on macOS, and adds the d3 decision models

Ollama, the most widely used way to run local models, shipped version 0.40.11 on 24 October with two additions that go beyond the usual bug-fix release. One gives it a new kind of output, speech. The other brings in a family of small models built for decisions rather than conversation.

Speech, on macOS for now

The headline line in the release is plain: “Ollama can now generate speech on macOS.” In practice that means ollama run can play text, or a chat model’s reply, aloud as the audio is generated, so a spoken answer streams out rather than waiting for the whole reply first. The example given is ollama run phonon "Hello from Ollama.", where “phonon” is the speech model. Applications can reach the same capability through an OpenAI-compatible /v1/audio/speech endpoint, which means software already written against that API can point at a local Ollama instead of a cloud service.

The honest limits: the release mentions only macOS, so this is a macOS capability today, and it does not say whether other platforms are planned. It also does not describe the “phonon” model, its size, or its licence, so if you intend to build on local text-to-speech, those are details to confirm before you rely on them.

The d3 decision models

The release also adds a family called d3, in 2B, 4B, 9B and 27B sizes. These are not general chat models. Per Ollama, d3 takes text, structured state, or an image and can “classify, answer yes or no, or score options with probabilities”, with an example of asking whether a customer message is a billing issue. That is the pattern of a small, fast model used for routing and gating inside a larger system, rather than for open-ended generation.

Two things to be clear about, because the release note is. First, these are not Ollama’s own models: the note says “d3 brings vLLM Semantic Router’s decision models to Ollama”, so they originate in the vLLM Semantic Router project. Second, the release page does not name a licence for them, so that too is unconfirmed from the source and worth checking before any commercial use.

Why it matters

On-device speech generation inside the runtime that most people already use lowers the barrier to building voice output that never leaves the machine, which is squarely the local-AI case: no audio sent to a cloud API, no per-character billing. The decision models point at a different and increasingly common use of small local models, as cheap, fast components that make a yes-or-no or classification call inside a pipeline, leaving the heavier generation to a larger model.

Both are genuine new capabilities rather than incremental tuning, which is why they are worth noting when most point releases are not. Ollama has been moving quickly on the Apple Silicon side in particular, having made MLX its default there earlier this month and rebuilt its Apple Silicon support over the summer. The caveats here are the undocumented pieces: a macOS-only speech feature, a speech model the release does not describe, and decision models ported from another project without a stated licence. Useful today, with a few details still to pin down.