Gemma 3 just released. In my setup I have a dedicated LLM server running Ollama as backend for the LLMs to serve to HA. I just updated Ollama to get Gemma 3 running. When I integrate Gemma 3 into HA, it is only working when I select „no control“ Gemma works inside HA and provides responses. When I select „assist“ I only receive an error „unexpected error during intent recognition“ according to the data sheet, Gemma 3 support function calling, so I guess it’s an issue on the HA side?
#Gemma 3 „Assist“ not working (Update: Now working)
forest canopy
solar iris
Yeah there will be an update somewhere, ollama or HA
whole gull
Seeing the same issue too with small thinker. Qwen keeps working and is the most reliable model so far, in terms of support from HA
forest canopy
That’s right. Qwen2.5 gave me the best experience + Mistral small 24b working very great as well. Gemma 3 on the other side looks very promising , after it also feature visual capabilities. So no need to have two models ruining like qwen2.5 + lama3.2 visual.
forest canopy
it’s an Ollama and/or Gemma and no HA issue when looking at this https://github.com/ollama/ollama/issues/9680
whole gull
it’s an Ollama and/or Gemma and no HA issue when looking at this https://github....
This is a really nice finding! I'm gonna give a try to phi4 instead to see how it works, since it supports function calling
forest canopy
There is now a variant of Gemma 3 which supports tool calling. I tested it in HA and it worked. https://ollama.com/PetrosStav/gemma3-tools:27b
whole gull
There is now a variant of Gemma 3 which supports tool calling. I tested it in HA...
How's the performance?
forest canopy
well it depends on your context size. what I really really like is how fast and good it perform with vision. Using "LLM Vision", I get almost instant result for analyzing pictures compared to lama3.2-12B Vision. The quality of recognition is way better compared to lama as well. The processing I do on my local LLM rig on my RTX3090s.
forest canopy
Gemma 3 „Assist“ not working (Update: Now working)
solar iris
I tried the PetrosStav model and it still doesn’t work for me. Qwen and others are fine. ##UPDATE: It appears its just the 12b model, the 4b and 27b work fine
woven plover
There is now a variant of Gemma 3 which supports tool calling. I tested it in HA...
I'm starting to feel like I just need to give up getting any of this work with any models smaller than +24b.