Is China’s Manus a Game Changer for Agentic AI?

New AI product releases—notably Alibaba’s QwQ-32B and Manus AI—debuting late last week show we’re entering a new era for the technology. 

These new technologies have important implications for government clients: AI is no longer just about capability—it’s increasingly about efficiency and autonomy. These two advancements signal a  shift in how AI is developed, accessed, and deployed, a change brought on by DeepSeek’s arrival in January 2025. The Chinese AI company demonstrated state-of-the-art performance can be achieved  without hyper-scale infrastructure, an action that challenged AI’s economic model.

Alibaba’s new model purports to build on DeepSeek’s efficiency breakthroughs to reduce compute demands. The company claims its new model enables  QwQ-32B to run on a single 24GB GPU  while competing with much larger models, such as the DeepSeek-R1. And  it has open-sourced its model weights—another move toward accelerating decentralized AI innovation.

On March 8,  Manus AI  introduced an  autonomous AI agent designed to independently complete multi-step digital workflows—researching, coding, and deploying applications  without human intervention. If verified, Manus AI’s advancement would mark a transition from  information retrieval  to  action execution and a step toward AI providing digital task automation instead of simply serving as chatbots.

Alibaba's Qwen team has introduced the QwQ-32B large language model (LLM) to compete with DeepSeek and similar lightweight reasoning models. Despite its smaller size, this model claims to rival the performance of larger LLMs on key benchmarks. Our team at Booz Allen has conducted an initial evaluation to understand its capabilities, limitations, and potential impact on the AI landscape. 

Qwen says the QwQ-32B model performs competitively on math, coding, and instruction-following benchmarks. Notably, this focus suggests that the model has been optimized for solving problems with verifiable answers at the expense of more creative or free-form responses.   

In our performance benchmarking, we found that despite its smaller size, QwQ-32B performed comparable to DeepSeek on evaluations such as “Google-Proof Question and Answer” (GPQA-Diamond) and the IFEval test. However, both models tended to randomly provide multilingual responses (e.g., mixing English words and characters with Chinese ones) that were otherwise coherent.   

In terms of robustness, initial red teaming tests indicate higher failure rates in areas such as toxicity, prompt injection, and data leakage compared to other models like DeepSeek-R1-Llama8B and Meta Llama8B.  

Dive deeper into our findings to understand the implications of QwQ-32B's performance on the AI market through the report below. Also, stay informed about the latest advancements and how they could impact your organization by subscribing to our newsletter.