Yes, you are right that for a model that wants to use tools, the environment was wrong. I didn't do that on purpose. I was simply interested in seeing what the answer to my question would be. The fact Gemma 4 wanted to use tools was a bit of a surprise to me - the Qwen model also can use tools, but it opted not to.

I think it is interesting to see, that when forced to derive the value on its own, Gemma gets it wrong while Qwen gets it right (although in a very costly way).

I also think that not using tools is better than hallucinating using them.

I'm not judging, just clarifying for others who might think that the model did actually run the tools (like I did initially).

Regardless of setup the LLM shouldn't hallucinate tool use.