Qwen Agent is Here

Following Qwen 2.5’s popularity among developers for AI agent development, Alibaba Cloud has now released Qwen Agent on GitHub.

Technically, Qwen Agent is a framework for developing LLM applications based on following instructions, tool usage, planning, and Qwen's memory capabilities. It also comes with example applications such as Browser Assistant, Code Interpreter, and Custom Assistant.

Qwen Agent provides atomic components, such as LLMs (which inherit from class BaseChatModel and come with function calling) and tools (which inherit from class BaseTool), along with high-level components like agents (derived from class Agent).

Furthermore, a user on X mentioned that Qwen Agent comes with a Gradio user interface.

Regarding accessibility, users can either utilise the model service by Alibaba Cloud’s DashScope or deploy and manage their own model service using the open-source Qwen models.

As per the README in the Github repository, “The LLM classes provide function-calling. Additionally, some Agent classes also are built upon the function calling capability, e.g., FnCallAgent and ReActChat.”

Almost a month ago, Qwen2-Math was released – the first series of mathematical LLMs of the Qwen family. 

Following that, it was upgraded and open-sourced to Qwen2.5-Math series, including base models Qwen2.5-Math-1.5B/7B/72B, instruction-tuned models Qwen2.5-Math-1.5B/7B/72B-Instruct, and mathematical reward model Qwen2.5-Math-RM-72B.

According to Qwen’s blog, the Qwen2-Math series only supports using Chain-of-Thought (CoT) to solve English math problems, while the Qwen2.5-Math series is expanded to support using both CoT and Tool-integrated Reasoning (TIR) to solve math problems in both Chinese and English. 

Developers Love Qwen

Notably, it seems like developers’ experience with Qwen has been largely positive. 

Beyond its ability to run on mid-range hardware, developers have been finding it better than popular LLMs like ChatGPT. According to one developer, since he started using Qwen 2.5 35B for coding tasks, he has not touched ChatGPT and only uses Claude for planning. 

“It is local, and it helps with debugging and generates good code. I do not have to deal with the limits on ChatGPT or Sonnet. I am also impressed with its ability to follow instructions and generate JSON output,” he further said. 

Another developer who extensively tested the model said he created a fully functional Pac-Man game in Python using the 72B model running locally in Q4 quantisation, complete with ghosts, playable map, and sprite loading functionality, outperforming Claude which only managed a basic map implementation. 

Qwen is also a reasonable choice for developers seeking to reduce dependency on cloud-based solutions, considering the $0.38 per million tokens they get compared to GPT-4o’s $5 per million tokens and Claude 3.5 Sonnet’s $3 per million tokens. 

Over the past few months, Amazon Web Services (AWS) has introduced Multi-Agent Orchestrator, a framework that offers a solution for managing multiple AI agents and handling complex conversations. 

The framework routes queries to the most suitable agent, maintains conversational context, and integrates with various environments, including AWS Lambda, local setups, and other cloud platforms.

Its GitHub repository highlights its capabilities with six specialised agents, including ones for travel, weather, math, and health. The orchestrator switches between agents to manage multi-turn conversations and diverse tasks while preserving context.

Meanwhile, Microsoft Research unveiled Magentic-One, a generalist multi-agent system capable of solving open-ended tasks across diverse domains. Available as an open-source tool on Microsoft AutoGen, Magentic-One helps developers and researchers create agentic applications for managing complex, multi-step tasks autonomously.

OpenAI introduced Swarm, a framework for building, orchestrating, and deploying multi-agent systems.

Similarly, IBM launched the Bee Agent Framework, an open-source toolkit for creating and deploying agent-based workflows at scale. Currently, in its alpha stage, Bee Agent supports various AI models and offers compatibility with IBM Granite and Llama 3.2 models.