I’m experiencing extreme frustration when trying to use the Gemini-2.0 model in production environments. The model consistently refuses to execute tool calls unless explicitly marked as mandatory. Moreover, it frequently instructs the user to manually trigger tools, e.g., “you should call the tool X with parameters Y,” instead of autonomously performing the action.
Interestingly, even smaller models handle tool calls effortlessly, unlike Gemini-2.0. I have tested this extensively through both LangChain and Vercel’s AI SDK, but the outcome remains unchanged—making Gemini essentially unusable in real-world scenarios.
Additionally, Gemini frequently encounters severe token repetition loops, rapidly exhausting resources. The lack of support, transparency, and effective resolution from the team makes it extremely challenging to adopt Gemini reliably.
This experience has been disappointing, to say the least. Urgent improvements in both functionality and support responsiveness are critically needed.