Qwen3 llm no_think

when I use my original and now your prompt the assistant just tells me it has trouble calling the tool. there must be something I’m missing or flipped off in the config for LLM Conversation. Qwen3 works fine through Ollama intregration minus the /think part.

I first used the default settings in Local LLM and did not run into any issues. I am running now with Max Tokens 1024 and Temperature of 0.1, TopK 40, Top P and Typical P at 1. which all might not be default. The temperature I would think matters as you don’t want it being too creative. I also have Prefer handling command locally turned on.

Hmm. I would check if you have any automations it might be conflicting with, I know I had to turn off a View Assist - Music Assistant Automation that it conflicted with.

Check the logs.

When I say “Play Hurt by Johnny Cash on the Basement Sonos”

I see this in the logs

2025-05-19 15:36:14.793 INFO (MainThread) [custom_components.llama_conversation.conversation] calling tool: {"name":"HassMediaPlay","arguments":{"name":"media_player.basement_sonos_3","media_content_type":"music","media_title":"Hurt","media_artist":"Johnny Cash"}}

When I say “Add Mow the lawn to my Home list in Google Tasks”

I see this in the logs:

2025-05-19 15:31:44.973 INFO (MainThread) [custom_components.llama_conversation.conversation] calling tool: {"name":"HassListAddItem","arguments":{"item":"Mow Lawn","name":"Home"}}

Usually if it says “I’m sorry! I encountered an error calling the tool. See the logs for more info.” - You might have to adjust what you are saying as it might be conflicting with something, and the logs usually give you a clue. I had an issue with using the word “to do” as it wanted to run the local “todo.get_item” function. So I switched to saying “task” or leaving that word out. Like the above sentence I didn’t say “Add Mow the lawn to my to do list” to avoid the issue.

Example: If I say “Play Master of Puppets by Metallica on Kitchen Sonos”
It gets mad at me. Why? I don’t have a Kitchen Sonos.

And the logs show:

2025-05-19 15:50:42.101 INFO (MainThread) [custom_components.llama_conversation.conversation] calling tool: {"name":"HassMediaUnpause","arguments":{"name":"media_player.kitchen_sonos"}}
2025-05-19 15:50:42.101 INFO (MainThread) [homeassistant.helpers.intent] Triggering intent handler <MediaUnpauseHandler - HassMediaUnpause>
2025-05-19 15:50:42.104 WARNING (MainThread) [homeassistant.helpers.service] Referenced entities media_player.kitchen_sonos are missing or not currently available

I’m no expert and this stuff changes everyday so I can’t help beyond that.

And if it matters I stay a little behind the current version and usually only add Minor updates without letters.

Local LLM Conversation integration is 0.3.8 (I think this is the latest)

So I am currently on 2025.05.1 (not the latest)

Good luck, wish I knew more.

It did not matter where or how often I put that no-think thing in the prompt. It just doesn’t work. Every time I get the response from the LLM the tags think and /think are there, though they are devoid of content. It’s also slow — I can see it’s taking a long time between the /think tag and the response.

I’m running the latest Ollama. My findings above hold true, whether I have set up the LLM to control the home or not.

See?

I’m using the Ollama integration.

Based on the screenshot, you’re using tool calling. Processing will be slow (large context size in request) if you’re not using top-tier GPUs.
As for tags, they will be removed in about a month in the new integration version; the Ollama developers added the necessary features just a week ago.

1 Like

It does not work with the Ollama integration. It works with Local LLM Conversation Integration.

https://community.home-assistant.io/t/local-llm-conversation-integration/675156

I am big fan of Ollama and use it everywhere else like LLM Vision and projects outside of HA, but it just pays no attention to the tag. Maybe in a future upgrade that will change.

Try Local LLM Conversation, it is fast and removes the thinking process.

For those interested, I made a fork of the extended OpenAI Conversation for HACS, which hides the s from the output. :slight_smile:

(How) can your fork be used with qwen3 running locally on Ollama?

The same way as in the original Extended OpenAI Conversation. :slight_smile:

Follow onboarding as usual, using Ollama’s OpenAI compatible endpoint.

Hey where do I get the API Key for Llama or LocalAI? Thank you very much :slight_smile:

Alright I got LocalAI and your Integration set up. But in the Chat there is still “ Okay, let’s see. The user is asking…”

Do I need to add the /no_think somewhere?