Alibaba has introduced Qwen3, the latest open-sourced large language model series generation.
The Qwen3 series includes six dense models and two Mixture-of-Experts (MoE) models, which aim to offer developers flexibility to build advanced applications across mobile devices, smart glasses, autonomous vehicles, and robotics.
All models in the Qwen3 family—spanning dense models with 0.6 billion to 32 billion parameters and MoE models with 30 billion (3 billion active) and 235 billion (22 billion active) parameters—are now open-sourced and accessible globally.
Qwen3 is Alibaba's first release of hybrid reasoning models. These models blend conventional large language model capabilities with more advanced and dynamic reasoning. Qwen3 can transition between "thinking mode" for complex multi-step tasks such as mathematics, coding, and logical deduction, and "non-thinking mode" for rapid, more general-purpose responses.
For developers using the Qwen3 API, the model provides control over the duration of its "thinking mode," which can extend up to 38,000 tokens. This is intended to enable a tailored balance between intelligence and computational efficiency. The Qwen3-235B-A22B MoE model is designed to lower deployment costs compared to other models in its class.
Qwen3 has been trained on a dataset comprising 36 trillion tokens, double the size of the dataset used to train its predecessor, Qwen2.5. Alibaba reports that this expanded training has improved reasoning, instruction following, tool use, and multilingual tasks.
Among Qwen3's features is support for 119 languages and dialects. The model is said to deliver high performance in translation and multilingual instruction-following.
Advanced agent integration is supported with native compatibility for the Model Context Protocol (MCP) and robust function-calling capabilities. These features place Qwen3 among open-source models targeting complex agent-based tasks.
Regarding benchmarking, Alibaba states that Qwen3 surpasses previous Qwen models—including QwQ in thinking mode and Qwen2.5 in non-thinking mode—on mathematics, coding, and logical reasoning tests.
The model also aims to provide more natural experiences in creative writing, role-playing, and multi-turn dialogue, supporting more engaging conversations.
Alibaba reports strong performance by Qwen3 models across several benchmarks, including AIME25 for mathematical reasoning, LiveCodeBench for coding proficiency, BFCL for tools and function-calling, and Arena-Hard for instruction-tuned large language models. The development of Qwen3's hybrid reasoning capacity involved a four-stage training process: long chain-of-thought cold start, reasoning-based reinforcement learning, thinking mode fusion, and general reinforcement learning.
Qwen3 models are now freely available on digital platforms including Hugging Face, Github, and ModelScope. An API is scheduled for release via Alibaba's Model Studio, the company's development platform for AI models. Qwen3 is also integrated into Alibaba's AI super assistant application, Quark.
The Qwen model family has attracted over 300 million downloads globally. Developers have produced over 100,000 derivative models based on Qwen on Hugging Face, which Alibaba claims ranks the series among the most widely adopted open-source AI models worldwide.
