Alibaba Launches Qwen3: New AI Models Rival Google and OpenAI

Chinese tech giant Alibaba has introduced Qwen3, a new family of AI models that the company says can compete with top models from Google and OpenAI, writes Techcrunch.com.

Qwen3 Models Show Strong Performance

The Qwen3 models, ranging from 0.6 billion to 235 billion parameters, are mostly available under an open license on platforms like Hugging Face and GitHub. Larger models with more parameters typically handle complex tasks better. The biggest model, Qwen3-235B-A22B, outperforms OpenAI’s o3-mini and Google’s Gemini 2.5 Pro on benchmarks like Codeforces (programming contests), AIME (math), and BFCL (reasoning). However, this model isn’t public yet. The largest public model, Qwen3-32B, still holds its own, beating OpenAI’s o1 on tests like LiveCodeBench for coding.

Advanced Features and Global Impact

Alibaba highlights Qwen3’s “hybrid” design, allowing it to reason through tough problems or respond quickly to simpler ones, with some models using a mixture of experts (MoE) architecture for better efficiency by breaking tasks into smaller parts handled by specialized sub-models. Trained on 36 trillion tokens—equivalent to 27 billion words—Qwen3 supports 119 languages and excels in tool-calling, following instructions, and handling data formats, with a dataset including textbooks, code, AI-generated content, and more. It’s also available via cloud providers like Fireworks AI and Hyperbolic. Tuhin Srivastava, CEO of AI cloud host Baseten, told TechCrunch that open models like Qwen3 show how global competition in AI is heating up, despite U.S. restrictions on chip sales to China, adding pressure on American AI labs to innovate faster.