Technology company Alibaba has released Qwen3, the newest version of its open-source large language model (LLM) family. The company designed the Qwen3 series to help developers create advanced applications for mobile devices, smart glasses, robots, autonomous vehicles, and more.
The Qwen3 line includes six dense models and two Mixture-of-Experts (MoE) models. These are now available for free worldwide. The dense models come in sizes ranging from 0.6 billion to 32 billion parameters. The MoE models include a 30 billion version with 3 billion active parameters, and a larger 235 billion version with 22 billion active parameters.
Qwen3 is Alibaba’s first model to use hybrid reasoning. This feature allows the model to switch between complex thinking for tasks like solving math problems or writing code, and quicker responses for simpler tasks. Developers using the Qwen3 API can also control how long the model stays in thinking mode, up to 38,000 tokens.
“Qwen3-235B-A22B significantly lowers deployment costs compared to other state-of-the-art models, reinforcing Alibaba’s commitment to accessible, high-performance AI,” the company said in a media release.
Step-by-step reasoning
Qwen3 was trained on 36 trillion tokens, double the amount used for Qwen2.5. This makes the model more capable in multiple languages, creative writing, tool usage, and step-by-step reasoning. It supports 119 languages and follows instructions in different languages more effectively.
Qwen3 supports agent-based tasks and includes strong function-calling abilities. It also uses a four-step training method that includes reinforcement learning and chain-of-thought techniques.
The models are available on Hugging Face, GitHub, and ModelScope. Developers can try them at chat.qwen.ai. API access will be offered soon through Alibaba’s Model Studio. Qwen3 also powers Alibaba’s AI assistant app, Quark.
Get the latest before it trends. Follow Back End News on LinkedIn, Facebook, X, YouTube, and TikTok for updates and in-depth coverage across the tech and security landscape.