Best cloud LLM for voice (and for free) | Home Assistant | Page 1

heady sky Jan 28, 2026, 12:23 PM

#

Hey guys, used the christmas break to build some ESP32 satellites to try out how good voice already works. As a base I used Homeway and their Sage AI as a free trial.

Now, at the end of the free trial and without being confident that integrating Homeway is a secure way, I want to replace it.

What is currently the best option for using cloud LLMs? In the best case for free and with good results in German...

Usecases are 95% connected to HA, triggering devices. But I guess the 5% share of other things might grow, when we get used to it.

I don't have a powerful GPU to work locally, but as I use Proxmox it could also be a way with ollama and a cloud model.

Thanks a lot

mossy basin Jan 28, 2026, 1:57 PM

#

Sorry, let me re-iterate. You want to have decent cloud LLM, that will work with tools and will be integrated to HA, and won't cost money to use. Right?

barren quartz Jan 28, 2026, 2:16 PM

#

#

Unfortunately, the 'best' cloud LLMS are not free.
And the 'free' LLMs are far from the best

sturdy wedge Jan 28, 2026, 2:25 PM

#

I’m using gpt-oss:120b-cloud via Ollama in free tier and yet to hit a limit. It’s great and has become my daily driver.

heady sky Jan 28, 2026, 2:39 PM

#

Well, was not expecting this kind of discussion, but yeah, if there are options available for free and the quality and speed is acceptable, happy to try them. For sure models which are for free will have limitations in speed and quality, but if I have to wait another second to get still a good answer, I am happy to wait.

barren quartz Jan 28, 2026, 2:40 PM

#

And honestly you don't need the best.

I use https://ollama.com/library/qwen3:4b-instruct-2507-q4_K_M

Which fits on a 6GB of VRAM GPU

qwen3:4b-instruct-2507-q4_K_M

Qwen3 is the latest generation of large language models in Qwen series, offering a comprehensive suite of dense and mixture-of-experts (MoE) models.

#

It is able to
Trigger media playback on both Speakers and TVs
Control stuff
Answer questions about the house
Call random scripts that are exposed

#

It can also do web searches with MCP, all of that in a 4b model

#

My point being, you don't need a 540B model from Gemini/ChatGPT to just control your home

sturdy wedge Jan 28, 2026, 2:45 PM

#

Ignore the conversation agent shown, it’s the gpt-oss model (just the way the integration was installed.

#Best cloud LLM for voice (and for free)