Alibaba unveils a new competitor for Google and OpenAI
Chinese tech giant Alibaba has launched its latest range of AI models under the name Qwen3. According to Alibaba, these models match – and in some cases would even exceed – the performance of Google’s best systems and OpenAI.
Most of the Qwen3 models are, or will soon be, available for download via an open license on AI platforms such as Hugging Face and GitHub. The models range in size from 600 million to 235 billion parameters. The number of parameters roughly indicates a model’s problem-solving capabilities; the more parameters, the better the performance usually is.
The rise of AI models from China, such as Qwen3, is putting pressure on more established players such as OpenAI to further improve their technologies. At the same time, Western policymakers have introduced restrictions to prevent Chinese companies from accessing advanced chips needed for AI training.
According to Alibaba, the Qwen3 models are hybrid models, meaning they can flexibly switch between quick answers and deep ‘reasoning’ about complex problems. This capability helps the models better control themselves, similar to the way OpenAI’s o3 models work, although it does result in slightly slower response times.
Some Qwen3 models use a ‘mixture of experts’ (MoE) architecture, where tasks are split and assigned to smaller, specialised sub-models. This increases efficiency in answering queries.
Qwen3 supports 119 languages, according to Alibaba, and was trained on a dataset of nearly 36 trillion tokens – an amount that includes textbooks, question-and-answer pairs, code snippets and AI-generated data.
Because of these improvements, the Qwen3 models perform significantly better than their predecessor Qwen2. While not clearly superior to recent toppers such as OpenAI’s o3 and o4-mini, they are among the strongest models today, according to tests.
Emerce
Subscribers 0
Fans 0
Followers 0
Followers