I’m using the same Vapi system prompt for OpenAI and Qwen, but Qwen either fails to call tools or calls them too early. My assistant is a restaurant booking agent with strict slot collection and confirmation rules. With OpenAI, the prompt works. With Qwen, it often: calls availability before all required slots are collected books before explicit confirmation repeats the flow misses tool calls in Chinese too For people who have deployed Qwen in Vapi: do you use a different prompt format from OpenAI? do you separate tool rules from tone/persona more aggressively? do you use shorter hard-gated instructions instead of long natural-language prompts? any model-specific settings or examples that improved tool timing?

Yeah, Qwen tends to be less strict with tool discipline than OpenAI. What’s worked for me is separating tool rules from persona completely, using short hard-gated instructions (step-by-step), and explicitly blocking tool calls until all required slots + confirmation are met. Also, adding negative instructions like DO NOT call tools unless…and simplifying the prompt (less natural language) helps a lot. Happy to take a quick look at your setup and tighten it for Qwen. @xwluvramen