Alibaba Qwen3
Alibaba launched Qwen3 as the new flagship generation of its Qwen family, building on the widely adopted Qwen2.5 line (released September 19, 2024). Qwen3 debuted Alibaba's first 'hybrid reasoning' models, able to switch between a deliberate 'thinking' mode for hard math, coding, and logic and a fast 'non-thinking' mode for general queries. The family spans six dense models (0.6B to 32B) and two MoE models (30B-A3B and the 235B-A22B flagship), trained on roughly 36 trillion tokens across 119 languages — nearly double Qwen2.5's 18 trillion. Alibaba claimed the flagship was competitive with frontier models like DeepSeek-R1 and Gemini 2.5 Pro, and that even its small 4B model could rival the older Qwen2.5-72B.
All eight Qwen3 models shipped as open weights under the permissive Apache 2.0 license, freely downloadable on Hugging Face, ModelScope, Kaggle, and GitHub, with a hosted option at chat.qwen.ai.
Qwen3 is best read as the maturing of what has become, by usage and derivative count, the most-adopted open-weight model family globally, anchoring China's position in open AI alongside DeepSeek. Unlike the Llama 4 episode, its launch claims were not marred by a benchmark-presentation scandal; the headline numbers were strong and broadly reproducible, though as always third-party real-world testing matters more than vendor charts. The genuinely useful innovation for buyers is the breadth of the lineup plus a true Apache 2.0 license, which lets organizations pick a size that fits their hardware and deploy commercially without restrictive terms. Executives should weigh provenance considerations given the model's Chinese origin and run their own evaluations on domain tasks rather than relying on aggregate leaderboard scores. Net, it is a credible, openly licensed alternative to Western frontier models, with the usual caveat that open benchmark leadership is fast-moving and short-lived.
Qwen3 signals that the leading open-weight models are increasingly coming from outside the U.S., giving executives a permissively licensed, deployable alternative — and a reason to evaluate models on their own tasks rather than trusting vendor benchmarks.