Ollama 0.40.9 adds speech with Phonon on Apple Silicon and d3 decision models, and brings Flint to Linux and Windows

Ollama 0.40.9 can now speak: on a Mac with Apple Silicon, it generates speech with a model called Phonon. The same release brings d3, vLLM Semantic Router's decision models , and Flint, the coding model that arrived in the previous release, now runs on Linux and Windows as well as macOS. The installers were uploaded on 23 October 2026 at 02:02 UTC , and the page's changelog line runs from v0.40.8 to v0.40.9.

What changed

Speech, on Apple Silicon. ollama run now plays text, or a chat model's reply, aloud as the audio is generated, and apps can stream audio through the OpenAI-compatible /v1/audio/speechendpoint. OpenAI-compatible means an app written for OpenAI's speech endpoint can send the same request to Ollama on your own machine. The release's example isollama run phonon "Hello from Ollama."

Phonon is Gradium's text-to-speech family: text goes in, spoken audio comes out. The default download,phonon:latest, is 104MB ; it is Phonon 31M English, with one voice, nameddefault. The 31M is its size in parameters, the numbers a model learned: 31 million. On an Apple M5 Max it takes 121 MB of memory once loaded. For another language you pick a 40M model — English, French, German, Spanish or Portuguese, or a multilingual one that covers all five. Every Phonon model produces 24 kHz mono audio, and the packages require Apple Silicon and Ollama 0.40.9 or later.

The licences are not the same. The 31M English checkpoint is released under Apache 2.0 , a permissive licence that allows commercial use; the 40M English, French, German, Spanish, Portuguese and multilingual checkpoints are released under CC BY-NC 4.0, which limits use to noncommercial purposes.

d3, a new family of decision models.d3 is a family of decision models from vLLM Semantic Router. You give it text, structured state or an image, and ask it to classify, answer yes or no, or score options with probabilities. It returns scores and probabilities for the options you provide, and does not write a conversational answer. The library page suggests it for routing support requests, triaging documents, prioritising cases or checking whether an image meets a criterion, and you can ask several questions about the same input in one request. The release's example isollama run d3 "A customer says their card was charged twice." --ask "Is this a billing issue?"

The idea is not new to Ollama; the family is. The library page's example sends its request to http://localhost:11434/v1/systemone, the endpoint of the decision models covered in our item onOllama 0.35. d3 itself requires Ollama 0.40.9 or later.

The family comes in 2B, 4B, 9B and 27B sizes. Plaind3runs the 4B model, a good starting point on a 16 GB machine. The library page pairs each size with the Apple Silicon system it was tested on and the memory it took once loaded:d3:2b on 8 GB, using 3.62 GB; d3 (4B) on 16 GB, using 5.63 GB; d3:9b on 32 GB, using 10.09 GB; and d3:27b, the most accurate in the publisher's own evaluation, on 64 GB, using 28.61 GB. Those are peak figures for the default MXFP8 packages on Apple Silicon with a 4,096-token context ; on other supported systems the same names select Q8_0 packages, and memory use will differ. MXFP8 and Q8_0 are two ways of storing the model's numbers in compressed form. The packages use a 4,096-token context , and Ollama supports text, structured state and still images for d3 — not video, and not Box/Track decisions. The models are released under Apache 2.0.

Flint on Linux and Windows.Flint now runs on Linux and Windows as well as macOS. Our item onOllama 0.40.8 explains what the model is and how much it asks of your machine.

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

What it means for you

If you have an Apple Silicon Mac and want a voice, start with ollama run phonon "Hello from Ollama."— the default model is a 104MB download. Feed it text written to be spoken: the library page warns that formatting marks such as Markdown headings may be read aloud. To save audio or call the speech API from your own app, the page points you to Ollama's speech guide. Two limits before you build on it. None of the five languages is Arabic. And only the 31M English model is under Apache 2.0 ; if what you are building earns money, the 40M models' noncommercial licence rules them out.

If you are on Windows or Linux,Phonon is not for you yet: its packages require Apple Silicon. Flint is: it now runs on your system , so the wait our 0.40.8 item advised is over. Check the model's size in that item before you pull it.

If you sort, route or score things, d3 is worth a try. On a 16 GB machine, start with plain d3; with 8 GB,d3:2bis the size tested on that much memory. On a Mac, that memory is the unified memory the processor and graphics cores share, so leave room for everything else you run. Keep each input short: the packages read 4,096 tokens at a time , a token being a word or a piece of one. And treat the numbers as the model's estimate: scores and probabilities are model outputs, not guaranteed real-world frequencies, so test it on examples from your own work before you set a threshold that triggers an action. Our guideWhich Machine Runs a Model Locally explains what decides whether a model fits.

If you connect coding tools or apps through Ollama,the fixes are reason enough to update: a streamed reply that fails halfway now says so instead of simply stopping , and coding tools started withollama launchget the right context length for cloud models.

How to update.The release's list of files carries an installer for each system, all uploaded on 23 October. On Windows, download and runOllamaSetup.exe; on a Mac,Ollama.dmg; on Linux, theinstall.shscript from the same list. Then open a terminal and typeollama -v: it prints the version you have installed, which should now be 0.40.9. If it shows an older number, quit Ollama completely, start it again and run it once more.

The source

The release page on GitHub, Ollama v0.40.9, its list of downloadable files with their upload times, and the two models' pages in Ollama's library, phonon and d3.