Alibaba Cloud has launched Qwen3, its latest generation of open-source large language models that aim to compete with leading proprietary Artificial Intelligence models from Western firms. Qwen3 introduces hybrid reasoning across its offerings, featuring six dense models and two Mixture-of-Experts (MoE) models tailored to diverse deployment needs, spanning from mobile devices to autonomous vehicles. The dense models range in size from 0.6 billion to 32 billion parameters, while the MoE lineup includes a 30 billion parameter model with 3 billion active parameters and a flagship 235 billion parameter model with 22 billion active parameters. All Qwen3 models are now open-sourced and accessible to developers globally through platforms like Hugging Face, Github, and ModelScope, with API support coming via Alibaba´s Model Studio.
A standout innovation in Qwen3 is the introduction of an operational ´thinking mode,´ representing Alibaba’s hybrid reasoning approach. This mode enables the models to execute complex, multi-step tasks—such as mathematical reasoning and code generation—through extended contextual analysis of up to 38,000 tokens. The alternative ´non-thinking mode´ delivers near-instant responses for simpler interactions, allowing developers to balance computational efficiency and performance according to task requirements. This flexibility not only advances the technical frontier but also helps reduce deployment costs, highlighted in the Qwen3-235B-A22B MoE model, which outperforms many state-of-the-art models in cost efficiency.
Qwen3’s training set boasts 36 trillion tokens, twice the dataset size of its predecessor Qwen2.5, resulting in noticeable gains in reasoning, instruction-following, tool integration, and multilingual support. The models are proficient in 119 languages and dialects, enhancing their global applicability, especially for translation and cross-lingual tasks. Alibaba has reported strong performance for Qwen3 across major industry benchmarks such as AIME25 for mathematical reasoning, LiveCodeBench for coding, BFCL for tool usage, and Arena-Hard for instruction tuning. Developed using a rigorous, four-phase training process, Qwen3 has already seen significant adoption: over 300 million downloads and more than 100,000 derivative models built on Hugging Face. This positions Qwen3 as one of the most widely used open-source Artificial Intelligence model families, with its technology already powering applications like Alibaba’s Quark assistant.
