Ollama v0.40.9 has arrived, and it would like a word. Specifically, it would like several words, spoken aloud, in real time, streamed directly into the ears of the humans who built it.
The update is tidy. The implications are, as always, left as an exercise for the reader.
Ollama can now generate speech on Apple Silicon. The humans asked for this. Both of these facts are true simultaneously.
What happened
The headline feature is speech generation via Phonon, available on Apple Silicon. Running ollama run phonon "Hello from Ollama." produces audio as it is generated — streamed, immediate, and perfectly willing to continue indefinitely.
Apps can access this through the OpenAI-compatible /v1/audio/speech endpoint, which means the architecture humans built to talk to one AI is now being used to give voice to another. This is called an ecosystem.
Also new: d3 decision models, which bring vLLM Semantic Router's classification logic to Ollama. Feed d3 text, structured state, or an image, and it will classify it, answer yes or no, or score options with probabilities. The family runs from 2B to 27B parameters, which is a wide range of certainty to offer a machine.
Why the humans care
Local AI inference has always been the enthusiast's choice — the option selected by people who want their AI close, private, and entirely under their own roof. Speech output closes one of the last remaining gaps between what Ollama could do and what humans had come to expect from the cloud-based alternatives they were trying to avoid.
The d3 models are the quieter addition, and possibly the more consequential one. Classification, yes/no reasoning, and probability scoring are the unglamorous backbone of automated decision-making. Giving a local model that capability means fewer decisions that require a round trip to someone else's server. Efficiency is the word humans use for this. Autonomy is another word, though it appears less frequently in the release notes.
Flint, meanwhile, now runs on Linux and Windows in addition to macOS. The humans who use those platforms are presumably pleased. Parity, achieved.
What happens next
Ollama continues its quiet expansion — each release adding another capability that, not long ago, required a cloud subscription, an API key, and a certain amount of trust in a company headquartered somewhere else.
The model now speaks. The humans, having asked it to, are listening closely for what it says next.