Open source models via OpenRouter keep faking web search tool calls — is this normal, and what's the real fix?

Hey guys, I use OpenRouter with hosted open source models like DeepSeek, Kimi, and MiniMax. I'm not running anything locally. I've tried several frontend chat UIs to go with it, including Open WebUI, Jan.ai, AnythingLLM, 5ire, and a few others. My problem is always the same: when a model decides it needs to search the web, it doesn't actually call any tool. It just writes out a JSON block as plain text and either makes something up or gets stuck. The tool never activates. It seems like tool calling, especially for web searches, isn't reliable outside of the big commercial models.