This model is no longer maintained. It has been replaced by ha-voice-granite-3b-GGUF, which is better on every axis that matters for Home Assistant voice control.

	this (deprecated)	replacement
Base	Qwen2.5-7B-Instruct	IBM Granite 4.1 3B
Quantization	Q4_K_M (~4.5 BPW)	Q8_0 (8.5 BPW)
Size	4.7 GB	3.6 GB
Latency	341 ms	240 ms
Context	8,192	65,536

Why you should switch

This model has two failure modes that were the reason for the replacement:

It frequently returns no speech after a tool call. Home Assistant then raises Last content in chat log is not an AssistantContent and the user hears "Unable to get response" — even though the action actually succeeded. Deployments had to patch around it with a synthetic "Done" response.
It goes silent entirely above ~1,200 prompt tokens — no content and no tool call. Real Home Assistant prompts are ~2,900 tokens, squarely inside the failing range. This is a capability limit, not a context limit: no num_ctx value fixes it.

The replacement was trained and validated specifically against both, and measures 0 missing responses in 72 tool-call round trips and 0 silent responses in 84 trials at production prompt size.

The files here are left in place so existing pulls do not break, but they will not be updated.

infohound/ha-voice-7b-GGUF

⚠️ DEPRECATED — use infohound/ha-voice-granite-3b-GGUF instead

Why you should switch

infohound/
ha-voice-7b-GGUF

⚠️ DEPRECATED — use `infohound/ha-voice-granite-3b-GGUF` instead