Alibaba unveiled Qwen3, a family of eight new AI models, on April 28, 2025, with its flagship 235 billion parameter Qwen3-235B-A22B model outperforming OpenAI's o1 and DeepSeek's R1 on several benchmarks, while approaching Google's Gemini 2.5 Pro. The models feature "hybrid reasoning" capabilities, allowing users to toggle between fast responses and more intensive thinking for complex problems, whilst supporting 119 languages.
The Qwen3 series includes two "mixture-of-experts" (MoE) models and six traditional "dense" models ranging from 0.6 billion to 235 billion parameters, all available under the Apache 2.0 open-source license. The models were trained on a dataset of 36 trillion tokens, double the size used for the previous Qwen2.5 version. Wei Sun, principal analyst of AI at Counterpoint Research, stated that Qwen3 represents a "significant breakthrough—not just for its best-in-class performance" but also for its "application potential." Ray Wang, an analyst specialising in US-China technology competition, noted that Qwen models have already generated over 300 million downloads worldwide and more than 100,000 derivative models on Hugging Face.
The Qwen3 models are now freely available for individual users on platforms like Hugging Face, GitHub, and others, as well as through Alibaba Cloud's web interface. Junyang Lin, a member of the Qwen team, indicated that future developments will include further scaling of model and dataset size, extending context lengths, broadening modality support, and enhancing reinforcement learning with environmental feedback mechanisms. The new model family offers particular advantages for enterprise users, as they can redirect OpenAI-compatible endpoints to the new model in hours, and on-premises running enables logging and inspection of all prompts and outputs.
Sources:
1.
Kyle Wiggers
2.
Dylan Butts
3.
Carl Franzen
