LM Studio and qwen3-8b - cant get i running...

[MarkusM73]

As described in a video by the author, I set up n8n-mcp in LM Studio. Health check and list all workflows are working fine. However, I am NOT getting a single workflow generated:

prompt: please, find a workflow template that will automatically backup my workflows to github

Error:
Model failed to generate a tool call

Failed to parse arguments for tool “search_templates”: params.task is not one of enum values: ai_automation,data_sync,webhook_processing,email_automation,slack_integration,data_transformation,file_processing,scheduling,api_integration,database_operations

{“name”: “search_templates”, “arguments”: {“searchMode”: “by_task”, ‘task’: “github_integration”}}

I have already tested various models and copied the prompt from GitHub, of course, but nothing helps. Does anyone have any ideas?

[MarkusM73]

great support…

[czlonkowski]

Hey @MarkusM73 — qwen3-8b is just a bit too weak for the tool-calling this MCP needs. It speaks the protocol but drops arguments and picks wrong tools on anything non-trivial, so "cant get it running" usually means it runs and then quietly makes a mess.

Upgrade and you will feel the difference:

• Qwen 3.5 9B — this is the sweet spot for most people. Dense, fits on a single consumer GPU, tool-calling is actually reliable, and it drives n8n-mcp well for real workflow building. Try this first.
• Qwen 3.5 27B — noticeably better if you can fit it. 262K context, predictable compute (no MoE routing), and it handles the partial-update diff flow without corrupting state. Worth it if you have the VRAM.

Couple of things to double-check while you are at it:

• LM Studio on a recent build — older versions had a broken tool-call streaming path that looks exactly like "stuck".
• Context window set to at least 32k.
• Start your flow from get_node_essentials, not get_node_info, to keep responses lean.

If it is still misbehaving after the upgrade, paste the LM Studio log and your prompt and I will dig in.