How to Fix Local Qwen Models Not Using Tools in Odysseus (Working Solution)

[tasfiqur-dev]

Enabling tool calling for Huihui-Qwen3-14B-Abliterated using llama.cpp in Odysseus

Environment

• OS: Windows 11
• Application: Odysseus
• Model: huihui-ai/Huihui-Qwen3-14B-abliterated-v2
• GGUF quantization: Q4_K_M
• Inference engine: llama.cpp server
• Model launcher: Odysseus Cookbook

The problem

I was trying to use my local Qwen3 model with Odysseus agent tools.

The model could generate normal responses, but when I asked it to inspect my workspace, it attempted to execute get_workspace as a Bash command instead of making a structured tool call.

I compared its behavior with a cloud model (DeepSeek), which could successfully use the same Odysseus workspace tools.

Investigation

I checked the llama.cpp /props endpoint and initially found:

"chat_format": "Content-only",
"chat_template_caps": {
  "supports_tools": false,
  "supports_tool_calls": false,
  "supports_object_arguments": false
}

However, the original Hugging Face model page provided a Jinja chat template that explicitly supports tool calling through:

• <tools> for tool definitions
• <tool_call> for function calls
• <tool_response> for tool results

This suggested that the model's tool-capable template was not being used correctly by the server.

Solution

I saved the model's original Jinja chat template as a separate .jinja file and configured llama.cpp through Odysseus Cookbook's Advanced → Extra args.

After correcting the launcher arguments and restarting the model server, I checked /props again.

The important values changed to:

"supports_tools": true,
"supports_tool_calls": true,
"supports_object_arguments": true,
"supports_parallel_tool_calls": true,
"supports_preserve_reasoning": true

The full original Jinja template is available on the Hugging Face model page:

https://huggingface.co/huihui-ai/Huihui-Qwen3-14B-abliterated-v2

Result

The local Qwen3 model could finally make structured tool calls through Odysseus.

I tested it by asking the model to inspect my selected workspace. It successfully retrieved the folder and subfolder structure.

There was one failed GREP operation during the process, but the model recovered automatically and still returned the folder structure.

Remaining observations

I noticed that the model sometimes repeated the same reasoning text multiple times in its thinking panel. I haven't investigated whether this is a model-generation issue or an Odysseus display/streaming issue.

Questions

1. Is there a recommended way to make Odysseus Cookbook automatically use the model's embedded tool-capable Jinja template?
2. Should tool-capable templates be detected automatically when launching GGUF models?
3. Is the repeated reasoning output a known issue with Qwen3 or the llama.cpp integration?

Hopefully this helps anyone trying to run local Qwen3 models as agents in Odysseus.