Chinese tech giant Alibaba has unveiled Qwen 3, a family of AI models that the company claims either match or exceed the capabilities of leading models from Google and OpenAI. The new models are available for download under an open license from platforms like Hugging Face and GitHub. They come in various sizes, ranging from 0.6 billion to 235 billion parameters. Generally, a higher number of parameters enhances a model's problem-solving abilities, but also increases computational demands. The emergence of Chinese AI models like Qwen 3 has escalated the competition in the global AI landscape, putting additional pressure on American labs such as OpenAI to develop more advanced technologies. This competition has also prompted policymakers to impose restrictions on Chinese AI companies, particularly in accessing the high-performance chips needed for training sophisticated models. Qwen 3 distinguishes itself by being a “hybrid” AI model, capable of balancing intricate reasoning tasks with quick responses to simpler queries. This dual mode allows the models to perform effective self-fact-checking, a feature akin to OpenAI’s models, though reasoning comes with higher latency costs. The Qwen team highlighted this balance by stating, “We have seamlessly integrated thinking and non-thinking modes, providing users with the flexibility to control the thinking budget.” Qwen 3 supports 119 languages and was trained on a vast dataset comprising nearly 36 trillion tokens. These tokens represent the raw data processed by the model, with 1 million tokens roughly equivalent to 750,000 words. The dataset includes a mix of textbooks, question-answer pairs, code snippets, and other content, which contributed to the model's enhanced performance. According to Alibaba, Qwen 3 significantly outperforms its predecessor, Qwen 2, in several benchmarks. Notably, the largest Qwen 3 model, Qwen-3-235B-A22B, has demonstrated superior performance over OpenAI’s o3-mini in programming contests on Codeforces, a platform for coding competitions. It also excelled in the latest version of AIME, a challenging math benchmark, and BFCL, a test assessing a model’s reasoning capabilities. However, this top-tier model is not currently available to the public. The largest public Qwen 3 model, Qwen3-32B, remains highly competitive even against proprietary and open AI models from rival labs. For instance, it outperforms OpenAI’s o1 model in the LiveBench accuracy benchmark and rivals DeepSeek’s R1 model. Qwen 3 boasts significant improvements in tool-calling capabilities, instruction-following, and data format replication. Beyond just releasing the models for download, Alibaba has made Qwen 3 available through cloud providers like Fireworks AI and Hyperbolic, broadening its accessibility. Tuhin Srivastava, co-founder and CEO of AI cloud host Baseten, commented on the significance of Qwen 3. “This is another strong indicator that open models are continuing to keep pace with closed-source systems like those developed by OpenAI,” Srivastava stated. Despite U.S. efforts to restrict chip sales and purchases involving China, state-of-the-art open models like Qwen 3 will likely see widespread domestic use. “It underscores the reality that businesses are leveraging both custom-built tools and off-the-shelf solutions from established companies like Anthropic and OpenAI,” he added. The release of Qwen 3 highlights the rapid advancements in AI technology and the growing importance of open-source models in the field. As these models continue to evolve, the global AI landscape is expected to become increasingly dynamic and competitive.