Model outputs json instead of actually calling the tool, what is the reason?

[JoaoHCopetti]

I'm using this model: https://ollama.com/library/qwen2.5-coder, 7b version.

But when I ask something it outputs a json instead of actually performing the task:

# Prompt
navigate to localhost:5173

# Output
{"name": "browser_mcp_browser_navigate", "arguments": {"url": "http://localhost:5173"}}

It calls the correct function, but doesn't perform anything.
Is it because the model size being small? Or is it related to something else?

Using cloud models works fine tho.

[bdougie]

This lacks some details that would be helpful, but I am assuming you are using a qwen model or similar that does has a non-standard tool calling set up.

Capabilities reference

Valid capabilities are tool_use and image_input; adding tool_use advertises native tool support, while omitting it (or using []) lets Continue rely on system message tools for Agent mode. [Model capabilities guide]

So, the practical suggestion:

If a model’s native tools are broken/incompatible and you want system message tools:
Configure the model without tool_use in capabilities (or set capabilities: []), and Agent mode will use system message tools automatically. [Qwen3-coder tools discussion; FAQs – tools not working]