Alibaba releases Qwen3: New open-source AI flagship model with hybrid reasoning

May 8, 2025

Alibaba introduces Qwen3, the latest generation of its open-source language model series, setting new standards in AI development.

The new model series comprises a total of eight variants: six lean models (with 0.6B, 1.7B, 4B, 8B, 14B and 32B parameters) and two Mixture-of-Experts (MoE) models (30B with 3B active parameters and 235B with 22B active parameters). All variants are now open source and available worldwide.

Hybrid reasoning for more flexible thought processes

With Qwen3, Alibaba is introducing hybrid reasoning models for the first time. These can dynamically switch between a ‘thinking mode’ for complex tasks such as mathematics, programming and logical reasoning, and a ‘non-thinking mode’ for quick, simple answers.

The API can also be used to control how long the model can remain in thinking mode, with a context window of up to 38,000 tokens. According to Alibaba, the new Qwen3-235B-A22B model (MoE) in particular scores highly with its high performance and significantly lower infrastructure costs compared to other models in its class.

Strong performance in languages, tools and reasoning

Qwen3 was trained on a dataset of 36 trillion tokens – twice as much as its predecessor, Qwen2.5 – and shows significant progress in multilingual capabilities, tool usage, reasoning and human-machine interaction. Highlights include:

  • Multilingual competence: The model supports 119 languages and dialects and delivers strong results in translation and multilingual query processing.
  • Agent capabilities: Qwen3 supports the Model Context Protocol (MCP) and powerful function calling, making it one of the leading open-source models for complex, agent-based applications.
  • Reasoning performance: In benchmarks for mathematics, programming and logical reasoning, Qwen3 outperforms all previous models in the series, including QwQ in thinking mode and Qwen2.5 in non-thinking mode.
  • Better human-machine interaction: Qwen3 delivers more natural creative text, more realistic role-playing and smoother multi-turn dialogue, significantly improving the user experience.

Thanks to a new model architecture, an expanded training dataset and optimised learning methods, Qwen3 achieves top scores in benchmarks such as AIME25 (mathematical reasoning), LiveCodeBench (programming competence), BFCL (tool and function usage) and Arena-Hard (instruction tuning). The development of the hybrid reasoning model is based on a four-stage training process: First, the model starts with a cold start, during which it learns to break down complex tasks into comprehensible steps (chain of thought). This is followed by reasoning-based reinforcement learning, the merging of thinking and non-thinking modes, and final fine-tuning through general reinforcement learning.

Free availability to promote innovation

The Qwen3 models are now available free of charge via Hugging Face, Github and ModelScope. Interactive tests can be carried out at chat.qwen.ai. API access will be provided shortly via Alibaba’s Model Studio. Qwen3 is also already being used in Alibaba’s in-house super assistant app, Quark.

Since its release, the Qwen model series has been downloaded more than 300 million times. Over 100,000 derivative Qwen models have been created by developers on Hugging Face, making Qwen one of the most widely used open-source AI model series worldwide.

Related Articles

Illegal cigarettes: Smugglers turn to drones and social media


New technologies are changing the black market for tobacco in Europe A recent report by the auditing and consulting firm KPMG shows that the illegal tobacco trade in Europe is increasingly being supported by digital means and modern transport technologies. Smugglers...

Share This