Apppublic

tuandunghcmut/qwen-agentworld-agent

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes

App README

🤖 Qwen-AgentWorld Tool-Calling Agent

A chatbot agent powered by `Qwen/Qwen-AgentWorld-35B-A3B` running on Hugging Face ZeroGPU.

The model acts as the agent policy: it decides which tool to call and with what arguments. The Space then executes the real tool and feeds the result back, looping until the model produces a final answer.

Tools available to the agent

Tool	What it does
`web_search`	DuckDuckGo web search (via `ddgs`)
`run_bash`	Runs a shell command inside the Space container
`run_python`	Executes a short Python snippet and returns its stdout
`write_todo`	Creates / overwrites the task to-do list
`update_todo`	Marks a to-do item done / not done

How it works

1.The user message is added to the conversation with the tool schemas (tokenizer.apply_chat_template(..., tools=TOOL_SPECS)).
2.The model generates; <tool_call>{...}</tool_call> blocks are parsed out.
3.Each tool runs, results are appended as role: tool messages.
4.Loop until the model answers in plain prose (max MAX_TOOL_STEPS).

Configuration (Space variables / secrets)

Variable	Default	Purpose
`MODEL_ID`	`Qwen/Qwen-AgentWorld-35B-A3B`	Model to load
`MAX_NEW_TOKENS`	`1536`	Generation length per step
`MAX_TOOL_STEPS`	`6`	Max tool-call iterations per turn
`GPU_DURATION`	`120`	ZeroGPU seconds reserved per call

Notes & caveats

—ZeroGPU requires a PRO account (or a Team/Enterprise org). This Space is hosted under a PRO account.
—The model is loaded in 4-bit (nf4) so the 35B MoE fits the ZeroGPU budget and loads quickly. Set the env var path / drop the quant config in app.py to run full precision on dedicated hardware.
—Qwen-AgentWorld is officially a world model (it can simulate environments). Here it is driven as a real tool-using agent; tool-call reliability may vary. Adjust the system prompt in app.py to taste.
—run_bash / run_python execute inside the Space sandbox. Keep the Space private if you don't want arbitrary code execution exposed publicly.