On Monday, Chinese tech giant Alibaba introduced Qwen 3, a groundbreaking family of AI models that the company asserts can match and, in some instances, surpass the leading models from Google and OpenAI.
Most of these models will soon be available for download under an “open” license on AI development platforms like Hugging Face and GitHub. The models vary in size, ranging from 0.6 billion to an impressive 235 billion parameters. In the world of AI, parameters are crucial as they correlate with a model’s problem-solving capabilities—generally, models with more parameters deliver superior performance.
The emergence of models like Qwen is intensifying competition, putting pressure on American labs like OpenAI to enhance their AI technologies. This competitive landscape has prompted policymakers to impose restrictions aimed at limiting Chinese AI companies’ access to essential chips for model training.
Alibaba describes its Qwen 3 models as “hybrid,” capable of both reasoning through complex problems and quickly addressing simpler requests. This reasoning ability allows the models to fact-check themselves, akin to OpenAI’s o3 models, though it may result in higher latency.
In a blog post, the Qwen team stated, “We have seamlessly integrated thinking and non-thinking modes, offering users the flexibility to control the thinking budget.” This innovative design allows users to configure task-specific budgets with greater ease.
Some Qwen 3 models utilize a mixture of experts (MoE) architecture, enhancing computational efficiency for query responses. MoE divides tasks into subtasks, delegating them to specialized “expert” models for optimal performance.
Supporting 119 languages, Qwen 3 was trained on a massive dataset of nearly 36 trillion tokens, which are the fundamental units of data processed by the models. To put this into perspective, 1 million tokens equate to approximately 750,000 words. The training data includes a diverse mix of textbooks, question-answer pairs, code snippets, and AI-generated content.
These enhancements significantly elevate Qwen 3’s capabilities compared to its predecessor, Qwen 2. While none of the Qwen 3 models are definitively superior to the latest offerings from OpenAI, such as o3 and o4-mini, they certainly stand out as strong performers.
On platforms like Codeforces, known for programming contests, the largest Qwen 3 model—Qwen-3-235B-A22B—has been shown to outperform OpenAI’s o3-mini and Google’s Gemini 2.5 Pro. Additionally, it excels in challenging benchmarks like AIME and BFCL, which assess a model’s reasoning abilities.
However, the Qwen-3-235B-A22B model is not yet publicly available.
The largest publicly accessible model, Qwen3-32B, remains competitive against several proprietary and open AI models, including DeepSeek’s R1. Notably, Qwen3-32B surpasses OpenAI’s o1 model in various tests, including the coding benchmark LiveCodeBench.
Alibaba claims that Qwen 3 “excels” in tool-calling capabilities and in following instructions, as well as replicating specific data formats. Alongside downloadable models, Qwen 3 is also accessible via cloud providers like Fireworks AI and Hyperbolic.
Tuhin Srivastava, co-founder and CEO of AI cloud host Baseten, remarked that Qwen 3 exemplifies the trend of open models keeping pace with closed-source systems like those from OpenAI. He stated, “The U.S. is doubling down on restricting sales of chips to China and purchases from China, but models like Qwen 3 that are state-of-the-art and open […] will undoubtedly be used domestically. It reflects the reality that businesses are both building their own tools and buying off the shelf from closed-model companies like Anthropic and OpenAI.”
