Introducing Qwen3 Language Model Family

Qwen3 LLM: New family excels in coding, math, and general tasks. Offers diverse models, multilingual support, hybrid thinking, and improved agent capabilities, all open-sourced with efficient performance gains.

Here are some insights from the provided document:

1. Thinking vs. Non-Thinking Modes: A User-Centric Design. Qwen3 introduces a unique hybrid approach, allowing users to choose between detailed, step-by-step reasoning ("Thinking Mode") and quick, near-instant responses ("Non-Thinking Mode"), offering unprecedented control over the model's cognitive process. This flexibility offers a better performance with efficiency.

2. True Multilingual Power: Beyond Simple Translation. Qwen3 supports 119 languages and dialects, showcasing a genuine commitment to multilingual capabilities, positioning it as a valuable tool for diverse international applications (not just English).

3. Massive Scale-Up in Training Data. Qwen3 was pretrained on approximately 36 trillion tokens, doubled that of Qwen2.5, collected from data web and PDF-like documents, including generating synthetic data by Qwen2.5-Math and Qwen2.5-Coder, which could be an interesting point to discuss the impact of high volume and quality data.

4. Accessibility is Key: Qwen3 models (including fine-tuned versions) are available on major platforms like Hugging Face, ModelScope, and Kaggle, this will encourage people to adopt and test it. Framework support (SGLang, vLLM) and local usage tools (Ollama, LMStudio) are also highlighted.

5. Agentic Tool Calling Prowess: Qwen3's coding and agentic capabilities have been significantly optimized, particularly in the realm of tool calling, with recommended use of "Qwen-Agent," indicating its strength in real-world task execution and orchestration.