If I ask Wappler to do something more complex, it starts giving me instructions on what to do....
"You can create the structure and general elements by calling the appropriate tool functions to create and edit files as needed.
{
“name”: “create_file_in_editor”,
“parameters”: {
“file_path”: “file:///D:/PRJ/Wappler/AITest/views/index.ejs”,
“file_type”: “content page”
}
}" etc....
I don't have experience with vLLM, but check their docs on tool calling which has a list of supported models with known issues and recommended flags.
Tool calling probably works fine on the LLM side. After all, Wappler even created a simple page for me: opened it, and entered new text into it. From this, I can conclude that tool calling works in principle. But what matters here is exactly how Wappler interacts with the LLM: what it sends to it and how it interprets the response. So far, I get the impression that the problem lies on Wappler’s side. It’s giving me errors even with Deepseek. It would be nice to be able to see a log of how Wappler operates in AI mode; then we could understand what’s going on. When there were issues with tool calling, Wappler used to throw errors; now it doesn’t, but its behavior is unclear—sometimes it creates a page, other times it tells me I should create one 
If Wappler has been tested with any LLMs that can be deployed locally, it would be great to see a list of them.
We use the openai SDK and the chat completions API for the custom endpoints.
I don't have a very good GPU in my machine but did a small test with LM Studio as local server and qwen/qwen3-4b-2507 as model. It does work fine with Wappler and tool calls do work.
thank you, i will try this tiny model