Alibaba Unveils ‘Hybrid’ AI Models Qwen3

The Chinese tech giant Alibaba has launched a new family of AI models, Qwen3, which are said to “match or even surpass in some cases” the best solutions from Google and OpenAI.

Introducing Qwen3!

We release and open-weight Qwen3, our latest large language models, including 2 MoE models and 6 dense models, ranging from 0.6B to 235B. Our flagship model, Qwen3-235B-A22B, achieves competitive results in benchmark evaluations of coding, math, general… pic.twitter.com/JWZkJeHWhC

— Qwen (@Alibaba_Qwen) April 28, 2025

Their size ranges from 600 million to 235 billion parameters. The models are “hybrid”—capable of both taking more time for reasoning and providing quick responses.

“We have seamlessly combined thinking and non-thinking modes, offering users flexibility […]. This design allows for easier budget adjustments for specific tasks,” the team noted in a blog post.

Qwen3 supports 119 languages and is trained on a dataset containing over 36 trillion tokens.

On the platform for evaluating programming skills, Qwen-3-235B-A22B outperformed o3-mini and Gemini 2.5 Pro. It surpassed o3-mini in the latest version of the AIME and BFCL math tests, which assess the ability to “reason” about problems.

Qwen-3-235B-A22B is not yet publicly available. Qwen3-32B is the largest among those open to the public. It exceeds o1 in several tests, including the LiveCodeBench programming benchmark.

Back in March, Alibaba introduced the reasoning-focused AI model QwQ-32.