The Qwen team at Alibaba has recently launched a new series of open-source AI models called Qwen3, which are making waves in the industry. These models include a total of eight new models, including two “mixture-of-experts” models and six dense models. The “mixture-of-experts” approach combines different specialty model types into one, activating only the relevant models when needed. This approach has been popularized by Mistral, a French AI startup.
The 235-billion parameter version of Qwen3, codenamed A22B, outperforms DeepSeek’s open-source R1 and even comes close to the performance of Google’s proprietary Gemini 2.5-Pro. This makes Qwen3 one of the most powerful publicly available models in the market.
One of the key features of the Qwen3 models is their “hybrid reasoning” capabilities, allowing users to switch between fast, accurate responses and more in-depth reasoning for complex queries. Users can toggle between “Thinking Mode” and “No Thinking Mode” to suit the task at hand.
These models can be accessed and deployed across various platforms, including Hugging Face, ModelScope, Kaggle, and GitHub. They are available under the Apache 2.0 open-source license, providing flexibility for users.
In terms of model training, Qwen3 represents a significant improvement over its predecessor, Qwen2.5. The pretraining dataset has doubled in size, and the training pipeline includes a three-stage pretraining process followed by a four-stage post-training refinement.
The Qwen3 models also offer expanded multilingual support, covering 119 languages and dialects. This makes them versatile for use in a wide range of linguistic contexts globally.
For enterprise decision-makers, the Qwen3 models offer a range of benefits, including easy integration with existing OpenAI-compatible endpoints, MoE checkpoints for GPT-4-class reasoning, and the ability to run weights on-premises for increased security.
Looking ahead, the Qwen team is focused on scaling data and model size further, extending context lengths, and enhancing reinforcement learning with environmental feedback mechanisms. The release of Qwen3 under an accessible license marks an important milestone in large-scale AI research, making state-of-the-art LLMs more accessible to researchers, developers, and organizations.
