Qwen3

Qwen3 is the latest large-scale language model launched by the Tongyi Qianwen team, aiming to provide users with efficient and flexible solutions through strong thinking and fast response capabilities. The model supports multiple thinking modes, can flexibly adjust the depth of reasoning according to task requirements, and supports 119 languages ​​and dialects, suitable for international applications. The release and open source of Qwen3 will greatly promote the research and development of large-scale basic models, helping researchers, developers and organizations around the world to build innovative solutions using cutting-edge models.

Demand population:

" Qwen3 is suitable for researchers, developers and organizations around the world, and is especially suitable for users who need to process multilingual data, complex inference tasks and fast response. Its flexible thinking mode and powerful multilingual support enable it to meet the needs of different fields, whether it is academic research, software development or commercial applications, to provide users with efficient and reliable solutions."

Example of usage scenarios:

Researchers can use Qwen3 's multilingual ability and powerful reasoning capabilities to conduct cross-language natural language processing research and promote the development of multilingual AI technology.

Developers can integrate Qwen3 into the software development process, use its thinking mode and fast response capabilities to provide users with a smarter interactive experience, such as intelligent customer service, code assistance, etc.

Enterprises can use Qwen3 's Agent capabilities to develop automation tools to improve work efficiency, such as automatic document processing, data analysis, etc.

Product Features:

Supports thinking mode and non-thinking mode. Users can flexibly adjust the inference depth of the model according to task needs. Complex problems can be gradually reasoned, and simple problems can be quickly responded to, achieving efficient and stable thinking budget control.

It supports 119 languages ​​and dialects, covering multiple languages ​​such as Indo-European language, Chinese-Tibetan language, Asian-African language and other languages, providing a wide range of possibilities for international applications, allowing global users to benefit from the powerful functions of the model.

The pre-trained data volume reaches 36 trillion tokens, covering 119 languages ​​and dialects. Through multi-stage pre-training, the model has strong language skills and general knowledge, and can effectively process longer inputs.

The post-training adopts a four-stage training process, including long thinking chain cold start, reinforcement learning, thinking mode fusion and general reinforcement learning, so that the model has the ability to think, reason and respond quickly.

Open source multiple model weights, including the large MoE model Qwen3 -235B-A22B and the small MoE model Qwen3 -30B-A3B, as well as multiple Dense models, are open sourced under the Apache 2.0 license, for user use and research.

Provides a variety of deployment options, including frameworks such as SGLang, vLLM, and local development tools such as Ollama and LMStudio, ensuring that users can easily integrate Qwen3 into their workflow.

It supports soft switching mechanism, and users can dynamically control the model's thinking mode in conversation, and switch the thinking mode round by round by adding /think and /no_think instructions to meet the needs of multiple rounds of conversations.

It provides Qwen-Agent tool, encapsulating tool call templates and parsers, reducing code complexity. Users can use MCP configuration files or custom tools to fully utilize the model's Agent capabilities.

Tutorials for use:

Visit the Qwen Chat web version or download the Tongyi APP, register and log in to your account.

Choose a suitable Qwen3 model, such as Qwen3 -30B-A3B or Qwen3 -235B-A22B, and choose a thinking mode or a non-thinking mode according to your needs.

Enter a question or directive in the dialog box and the model will respond according to the set pattern. For complex problems, the model will gradually reason and give answers; for simple problems, the model will respond quickly.

If you need to use a model for development or deployment, you can refer to the official documents and use the model weights on platforms such as Hugging Face and ModelScope, and integrate them with frameworks such as SGLang and vLLM.

For local development, you can use tools such as Ollama and LMStudio to run the model and interact with the model through the command line or graphical interface.

In multiple rounds of conversations, you can dynamically switch the thinking mode of the model to meet different conversation needs by adding /think and /no_think instructions to user prompts or system messages.

To use Qwen-Agent's Agent capabilities, you can define a tool list, use MCP configuration files or custom tools to make tool calls through Qwen-Agent encapsulated templates and parsers.