Global Versatility Supporting 119 Languages and Dialects
Enhanced Programming and Agent Capabilities
April 29, 2025: Alibaba Qwen team officially announced next-generation large language model Qwen3. Under the slogan "Think Deeper, Act Faster," this model incorporates large-scale technical innovations differentiating it from existing AI models. Three notable points: (1) "Hybrid thinking system" allowing users to adjust AI thinking depth themselves; (2) Global versatility supporting 119 languages and dialects; (3) Significantly enhanced programming and agent functions. The hybrid thinking mode: Qwen3 most notable feature is the "hybrid thinking mode." Users can now choose whether AI provides answers through deep thinking or quick response depending on the situation -- "Thinking Mode" (slower, deeper reasoning for complex problems) vs "Non-Thinking Mode" (faster, more direct responses for simple queries); users can also set a "thinking budget" controlling how much computation the model spends reasoning before answering. Why this matters: most AI models are either always-on reasoning models (slow, expensive) or non-reasoning models (fast, less accurate for complex tasks); Qwen3 hybrid approach lets users optimize the cost-quality tradeoff for each specific query; this is particularly valuable for agentic applications where some subtasks require deep reasoning while others need speed. Performance benchmarks: Qwen3-235B (largest model) approaches or exceeds GPT-4o and Claude 3.7 Sonnet on major benchmarks including MATH, MMLU, and coding tasks; the 30B parameter model achieves competitive performance with models 5-10x larger, demonstrating improved training efficiency. Open source release: Qwen3 models released on Hugging Face under Apache 2.0 license, enabling commercial use without restrictions -- a deliberate strategy to accelerate ecosystem adoption and compete with Meta Llama for developer mindshare.
