BUG - Tool calling not working with Ollama 0.9.2 + Qwen3 models (4B / 32B)

[tesla-yan]

Hi, thanks for the excellent project.

When using deepagents together with Ollama 0.9.2 and Qwen3 series models, I found that tool calling does not work at all.

Environment:
Ollama 0.9.2
Models tested: qwen3:4b, qwen3:32b
deepagents (latest)

Logging:

Optional[{
	messages=[
	UserMessage{content='"what is langgraph?"
', metadata={messageType=USER}, messageType=USER}
	AssistantMessage [messageType=ASSISTANT, toolCalls=[], textContent={"name": "internet_search", "parameters": {"q": "langgraph", "limit": "10", "raw": "true"}}, metadata={role=ASSISTANT, messageType=ASSISTANT, finishReason=STOP, refusal=, index=0, annotations=[{}], id=chatcmpl-322}]
	]
	files={}
	}]

Thanks for taking a look!

[tesla-yan]

@bsorrentino

[tesla-yan]

@bsorrentino

By the way.

I tested both the OpenAI Chat API style and the Ollama Chat API.
The spring-ai-agent example here —
https://github.com/langgraph4j/langgraph4j/tree/main/spring-ai/spring-ai-agent/src/test/java/org/bsc/langgraph4j/spring/ai/agentexecutor
—
can successfully perform tool calls with a Qwen3 model on Ollama 0.9.2.

However, langgraph4j-deepagents cannot perform tool calls under the same setup.

I’m trying to understand why the spring-ai-agent test works, but deepagents does not support tool calling in this scenario.

[bsorrentino]

Hi @tesla-yan I'm verifying this issue.

Deepagents use complex prompt and seems to me that QWEN doesn't understand that it should involve sub agents as tools, while gpt-4o-mini works as expected

[bsorrentino]

Hi @tesla-yan I got good result from qwen3:14b specifying a more complex question like :

I want that you perform a deep research on:
 
LangGraph4j project and in particular
 - What is langgraph4j project?
 - What about its integration with Spring AI ?
 - What about Workflow design pattern ?
 - What are the feasible use case scenarios ?
 - What about its tools ecosystem ?

[tesla-yan]

Hi @tesla-yan I'm verifying this issue.

Deepagents use complex prompt and seems to me that QWEN doesn't understand that it should involve sub agents as tools, while works as expectedgpt-4o-mini

Yes, it also works when I use a Llama model. I’m not yet sure whether the issue is coming from the model itself.

[tesla-yan]

Hi @tesla-yan I got good result from qwen3:14b specifying a more complex question like :

Thanks! I’m not fully sure I understand — do you mean that tool calling is already working when using qwen3:14b with more complex prompts?

[bsorrentino]

Hi @tesla-yan I got good result from qwen3:14b specifying a more complex question like :

Thanks! I’m not fully sure I understand — do you mean that tool calling is already working when using qwen3:14b with more complex prompts?

I try explain better.

Deepagents use complex system prompts to describe their behaviour, for this reason we need models that must understand them

• I've seen that less powerful models do not call subagents (i.e. tools) because don't take in consideration starting a deep process.
• I've tried to specify better user query to force model to start the deep process

Summarising: If you want use local (less powerful ) model you must working on the prompts and query, it is not problem related to implementation