Ollama v0.40.11: Speech on macOS and d3 Models

4 min read · Last updated: October 25, 2026

Ollama v0.40.11, released on 24 October, adds two things: speech generation on macOS, and the d3 family of decision models. ollama run can now read text or a chat model's reply aloud as the audio is generated, and apps can stream audio through the OpenAI-compatible /v1/audio/speech endpoint.

Came here from Instagram, LinkedIn, X or YouTube? Here is the full source list you came for. Everything below is built from these, nothing else.

The release

• Ollama v0.40.11 release notes on GitHub, published 24 October 2026

Ollama, the tool for running open models on your own machine, has not shipped a long changelog this time. The release notes for v0.40.11 hold two features, one new contributor and one commit since the release candidate. That is a short list, so I will go through each item, show the command that goes with it, and be clear about what the notes do not say.

Key takeaways

• Ollama v0.40.11 was released on 24 October and is marked as the latest release on GitHub.
• Speech is new, and it works on macOS. ollama run plays text or a chat model's reply aloud while the audio is still being generated.
• Apps can stream audio through the OpenAI-compatible /v1/audio/speech endpoint.
• d3 brings vLLM Semantic Router's decision models to Ollama in four sizes: 2B, 4B, 9B and 27B.
• d3 takes text, structured state or an image, and can classify, answer yes or no, or score options with probabilities.

What is new in Ollama v0.40.11?

Ollama v0.40.11 adds speech generation on macOS and the d3 decision models. It also welcomes one first-time contributor, @SynapShift, whose first contribution is pull request #19173. The release is one commit on main since the previous tag, and GitHub shows it as signed with a verified signature.

Change	What it does	Where you use it
Speech	Generates audio from text, on macOS	ollama run and /v1/audio/speech
d3 decision models	Classify, answer yes or no, or score options	ollama run d3 with --ask
New contributor	@SynapShift, first contribution in #19173	Not user-facing

How does speech work in Ollama on macOS?

Ollama can now generate speech on macOS. ollama run plays text, or a chat model's reply, aloud as the audio is generated. You do not wait for a finished audio file before you hear anything. The notes name one example model, phonon, and show a single command.

Before this release, ollama run printed text. After it, the same command can talk.

ollama run phonon "Hello from Ollama."

The release notes say it plays the reply aloud as the audio is generated. That is the behaviour change: playback starts while generation is still running.

Two limits are worth stating plainly. The announcement says speech generation is for macOS. It does not say anything about other operating systems, so do not assume it works there. It also names only one speech model in its example, so I will not list others.

Can my app stream audio from Ollama?

Yes. Apps can stream audio through the /v1/audio/speech endpoint, which Ollama describes as OpenAI-compatible. In plain terms, code written against that style of speech API can point at Ollama instead of a hosted service, and receive audio as it is generated.

The release notes do not include a request body, a list of voices, or the audio formats the endpoint returns. I am not going to guess them. The notes link to Ollama's speech documentation, so check the field names there before you write the client.

This is the part that matters for projects. The ollama run change is for you at a terminal. The endpoint is for your product.

What are the d3 decision models?

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

That is a different job from chat. A chat model writes a paragraph. A decision model returns a verdict, which is easier for code to act on.

Size	Listed in the release notes
2B	Yes
4B	Yes
9B	Yes
27B	Yes

The notes do not say how much memory each size needs or which one to pick, so I will not either.

How do you run d3 from the command line?

Pass the input text, then use the --ask flag to put your question to the model. The example in the release notes gives d3 a customer message and asks whether it is a billing issue.

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

The notes describe three kinds of answer:

1. Classify. Put the input into a category.
2. Yes or no. Answer a direct question, as in the example above.
3. Score options. Rank choices with probabilities.

The notes do not show the exact output format for any of the three. Run the example once and look at what comes back before you build on it.

What should you check before upgrading to Ollama v0.40.11?

The release notes list no breaking changes, no removed flags and no migration steps. They also do not state a minimum macOS version. That is all the notes cover, so run your usual checks.

1. Read the release page once, so you see the same notes I did.
2. Update on a non-critical machine first if Ollama sits behind a service you depend on.
3. If you want speech, run the phonon example on a Mac and confirm you hear the audio.
4. If you want d3, run the billing example, then try one of your own inputs.
5. If you plan to use /v1/audio/speech, read Ollama's speech documentation before writing client code.

There is no data migration in the notes, so nothing here suggests you need to back anything up. That is an absence of information, not a promise.

What does this mean for your projects?

There are two different things to take from this release.

Speech moves text-to-speech toward the machine you already use. If you build a local assistant, a reading tool or an accessibility feature, an OpenAI-compatible speech endpoint means less new client code. Streaming means the first words can play before the last ones exist. Keep the macOS limit in mind if your users are not all on Macs.

d3 turns a model into a decision step. Routing a support ticket, flagging a billing complaint or choosing between options are classification jobs. A model that returns a yes or no, or probabilities for each option, fits that shape better than parsing free text.

If you are new to running models locally, our guide to Google Gemma 4 and running free open-source AI locally covers the basics. For another voice model release, see the Block Dessa voice model. If you swap models often, read about AI model swapping.

FAQ

What is new in Ollama v0.40.11?

Two features: speech generation on macOS and the d3 decision models. The release notes also credit one new contributor, @SynapShift. GitHub shows the release as published on 24 October.

Does Ollama speech work on Windows or Linux?

The release notes say Ollama can now generate speech on macOS. They do not mention Windows or Linux. Treat speech as macOS-only until Ollama's documentation says otherwise.

What is the Ollama /v1/audio/speech endpoint?

It is the endpoint apps use to stream generated audio from Ollama. The release notes call it OpenAI-compatible. They do not list its parameters, so check Ollama's speech documentation for the request format.

What can the d3 models do?

You give d3 text, structured state or an image. It can classify the input, answer yes or no, or score options with probabilities. It comes in 2B, 4B, 9B and 27B sizes and brings vLLM Semantic Router's decision models to Ollama.

Try the speech example on a Mac today, then browse the free tools on Art of Code while the model downloads. If you build with local models, the Gemma 4 local setup guide is the next read.

Written by Adithya Guttha, Founder of Art of Code.