【Candy Apples Archives】

On April 29,Candy Apples Archives Alibaba unveiled Qwen3, its newest large language model and China’s first hybrid reasoning model that integrates both fast and slow thinking modes to reduce computational costs.
The Qwen3 series includes a range of models, such as the fine-tuned Qwen3-30B-A3B and its pre-trained base, now available across major platforms. Alibaba Cloud also open-sourced two Mixture-of-Experts (MoE) models: the flagship Qwen3-235B-A22B, with over 235 billion parameters, and the lightweight Qwen3-30B-A3B, with 30 billion total and 3 billion active parameters. According to Alibaba Cloud, Qwen3-235B-A22B delivers competitive results in coding, math, and general reasoning benchmarks, rivaling top models like DeepSeek-R1, 01.AI’s o1 and o3-mini, Grok-3, and Gemini 2.5 Pro.
Qwen3’s two reasoning modes allow users to toggle between in-depth step-by-step answers or rapid responses, depending on task complexity, a flexible design aimed at balancing speed and intelligence. [Alibaba, in Chinese] 

(Editor: {typename type="name"/})