Mar 17, 2025

Amanda Liang, Taipei; Elaine Chen, DIGITIMES Asia Monday 17 March 2025 Credit: AFP After DeepSeek sparked a revolution in China's AI industry in early 2025, Alibaba's Tongyi Qianwen QwQ-32B is poised to become the next widely adopted large model, thanks to its parameters and open-source advantages. While DeepSeek-R1 brought large models into everyday conversations, QwQ-32B is expected to take them further, embedding them in practical, real-world applications. Though both models deliver comparable performance, QwQ-32B demonstrates a broader scope of adaptability in real-world use. From enterprise-level solutions to personal development tools, and cloud to local deployments, QwQ-32B provides a competitive edge at an extremely low cost. The shift from DeepSeek-R1 to QwQ-32B signifies a dramatic reduction in the computational power required for top-tier performance models. This change is expected to disrupt the tech giants who have long relied on costly computing resources. Previously, an Apple Mac Studio with 512GB of memory, costing nearly CNY100,000 (approx. US$13,816), was necessary to run the full version of DeepSeek-R1. However, now, a Mac mini costing only a few thousand CNY can run QwQ-32B and provide nearly the same experience. Moreover, QwQ-32B's smaller parameter model has a natural advantage in inference speed under the same hardware conditions, offering faster response times and enhanced parallel processing capabilities. For small and medium-sized teams, startups, and individual developers, this significantly lowers the barrier to deploying inference models. With its lightweight architecture of 32 billion parameters, QwQ-32B has carved out a niche in China's AI sector. QwQ-32B has gained favor from various domestic AI chip platforms, including vendors like Sophgo and Biren Technology. As these platforms deeply integrate with Tongyi Qianwen's large model, China's AI industry is beginning to break free from the shackles of computational power constraints. This could mark the beginning of China's AI sector asserting its influence on the global stage. As an open-source large language model, Tongyi Qianwen has become a favorite within the developer community, thanks to its flexible customization features that allow developers to adjust and optimize the model to suit their specific needs. This makes it highly suitable for scientific research and technical development. The model has not only been adopted and deployed by several well-known overseas platforms but has also remained at the top of the global AI open-source community's trend rankings on Hugging Face, making it one of the most popular open-source large models worldwide. Additionally, Qwen-derived models have surpassed 100,000, overtaking Meta's Llama model. This cross-border influence indicates that Tongyi Qianwen is poised to make its mark on the global stage. From DeepSeek-R1 to QwQ-32B, China's large models are shifting the AI industry's focus from a parameter race to application precision. The question remains: will these developments shake up the future technological direction of AI giants in Europe and the US? And can China's AI chipmakers, including Ascend and Biren, capture more of the market share currently dominated by Nvidia? Behind this technological revolution lies a potential restructuring of the current order. The emergence of Tongyi Qianwen QwQ-32B could lead to significant disruptions in the European and American AI industries, as the open-source LLMs born out of China's computing power constraints begin to reshape the landscape. Article edited by Joseph Chen Related stories