# Ollama 0.40.9 adds local text-to-speech on Apple Silicon, d3 decision models, and Flint on Linux and Windows

> OnPrem World · https://onpremworld.com/en/news/2026-10-ollama-0-40-9-speech-d3 · CC BY 4.0

Published: 2026-10-22 · Runtimes & apps

Ollama v0.40.9, released on October 22, can generate speech with the Phonon model on Apple Silicon: `ollama run` reads text or a chat model's reply aloud as the audio is generated, and apps can stream audio from the OpenAI-compatible `/v1/audio/speech` endpoint. The default Phonon model uses Gradium's 31M-parameter English checkpoint, takes up to 4,096 characters per request and returns 24 kHz mono WAV or PCM. The release also adds d3, vLLM Semantic Router's Apache 2.0 decision models in 2B, 4B, 9B and 27B sizes, which classify, answer yes or no, or score options with probabilities from text, structured state or a still image. Recursive's Flint coding model, macOS-only in 0.40.8, now also runs on Linux and Windows.

## Why it matters locally

Voice output and yes-or-no routing decisions used to mean calling a cloud API; now both run on the same local server you already use for chat.

## For the market

Speech through an OpenAI-compatible endpoint means apps written for cloud text-to-speech could point at a Mac on the desk instead, which likely widens Ollama's role from a chat runner to a general local AI server. Small decision models like d3 also suggest a trend toward routing each request to the cheapest model that can handle it.

## In plain words

Ollama, a free app for running AI on your own computer, can now read answers out loud on Apple Silicon Macs, without sending your text anywhere. It also gets a new family of small models that answer simple questions like "is this a billing complaint?" with a yes, a no or a score.

## For practitioners

Phonon's default checkpoint is English only and accepts `wav` or `pcm` at 24 kHz; other formats, speeds or `instructions` return a 400 error, so check your client before swapping it in for a cloud voice. Ollama suggests 8, 16, 32 and 64 GB of system memory for the d3 2B, 4B, 9B and 27B packages, which run with a 4,096-token context, so test them on your own labelled examples before putting one in front of a larger model as a router.

## Source

[Ollama v0.40.9 release notes (GitHub)](https://github.com/ollama/ollama/releases/tag/v0.40.9)

## On this site

- Hardware: [M4 Pro](https://onpremworld.com/en/hardware/apple-m4-pro.md)
- Hardware: [M5](https://onpremworld.com/en/hardware/apple-m5.md)
- Hardware: [M5 Pro](https://onpremworld.com/en/hardware/apple-m5-pro.md)

[All news](https://onpremworld.com/en/news.md)

Summaries are written by OnPrem World from the official source. We link to the original and never republish it.