Ollama 0.40.9 added local speech and a new family of decision models. Here’s how I’d set up a safe test of both on a spare machine or a laptop. It’s a plan, not a benchmark, so adjust it to your hardware.
1. Update and check
Update Ollama using whatever method you originally installed it with, then confirm the version:
ollama --versionYou want 0.40.9 or later. Keep the server bound to localhost, which is the default, and don’t expose it to your network unless you’ve put authentication in front of it.
2. Speech with Phonon (Apple Silicon only for now)
According to the release notes, Phonon speech currently needs Apple Silicon. This is the example from the notes:
ollama run phonon "Hello from Ollama."The text plays aloud as the audio is generated. Apps can also stream audio through Ollama’s OpenAI-compatible /v1/audio/speech endpoint, so anything that already speaks that API shape can point at your local server instead. Check Ollama’s docs for the exact request format before you wire anything up.
3. d3 decision models
d3 models classify, answer yes or no, or score options with probabilities. They come in 2B, 4B, 9B and 27B sizes, so start with the smallest one that fits your RAM or VRAM. The example from the release notes:
ollama run d3 "A customer says their card was charged twice." --ask "Is this a billing issue?"Good test ideas: ticket triage, “does this log line look like an error?”, or “is this message safe to auto-reply to?” Feed it a handful of examples and see whether the probabilities match your own judgment before trusting it with anything.
4. Be careful with new models
Review before you install. Phonon and d3 are brand-new models. Before you pull one, read its page on ollama.com, check the size, the license and who published it, and make sure you have the disk space. Test on a machine without sensitive data first.
5. Flint on Linux and Windows
If you were waiting to try the Flint coding model from 0.40.8, it now runs on Linux and Windows too. Same advice applies: read the model page first.
6. Clean up
ollama list
ollama rm <model-name>Remove anything you’re not going to use so you’re not carrying around gigabytes of models you tested once.
Have a blessed day! And Happy Hacking!
