Alibaba has announced the launch of a new family of large language models (LLM) called Qwen3, which includes models with parameter counts ranging from 0.6 to 235 billion. Company representatives assure that in key AI tasks, these models demonstrate results comparable to or better than solutions from giants like OpenAI and Google.
This is reported by Business • Media
Features and Technical Specifications of Qwen3
The new models are distributed under an open license and are already available on platforms Hugging Face and GitHub. They support a hybrid processing mode that allows the models to perform both simple and complex computational operations, optimizing resource usage.
Some versions are built on the MoE (Mixture of Experts) architecture, which distributes tasks among specialized sub-models. According to company representatives, the volume of training data amounted to nearly 36 trillion tokens, including training materials, code, questions and answers, as well as synthetic data.
Testing Results and Development Prospects
According to tests on the Codeforces and AIME platforms, the largest Qwen3 model surpassed the performance of OpenAI’s o3-mini and Google’s Gemini 2.5 Pro. However, the version with 235 billion parameters is not yet available, but the Qwen3-32B model is already actively used in the market, demonstrating better results in a number of coding benchmarks compared to OpenAI’s o1 model.
According to company information, Qwen3 models can already be launched through cloud platforms Fireworks AI and Hyperbolic. Experts believe that despite export restrictions to other countries, Chinese AI developers are increasingly competing with Western industry leaders.