Pre-release of v3.46.0, for field validation of the bundle before the stable cut. HACS offers it only to users who opted into betas. It contains everything on main as of 41ce723, so it includes all of v3.46.0-beta.4.
New since beta.4
- A sleeping Ollama server ahead of a fallback costs a conversation turn one 8-second wait, not one per model call (#1027, reported on #1022). A turn that uses tools makes several model calls, and each one waited 8 seconds for the server that was not there before moving to the fallback. Now, once a call in a turn cannot connect to a server, the rest of that turn tries the other servers first. The unreachable one is tried again only if they all fail, and the next turn starts fresh. This covers a server that cannot be connected to at all; one that connects and then fails is retried as before.
- The docs now say that every Ollama server in a fallback chain needs the model pulled under the same name, tag included.
What to try
- With the primary Ollama server asleep or off and a fallback configured, ask for something that needs a tool (turn a light on, read a sensor). The reply should take about 8 seconds longer than normal, not 8 seconds per step.
- Wake the primary server and ask again: the next turn should use it.
Still being validated from earlier betas: remembering turns trimmed from a long conversation (#1048), fewer closing questions on voice satellites (#954), summarizing a conversation that used tools on OpenAI or Anthropic (#1045), a second Ollama server as a fallback running the feature's model (#1022), speaker recognition on voice satellites, memory management, and Top P per feature.
Tested against Home Assistant 2026.11.2. The new behaviour is covered by automated tests but has not yet been run against a real sleeping server.
The full notes are in the [Unreleased] section of CHANGELOG.md.