Alibaba has recently released the new Qwen3 series open-source large language model lineup, and it’s a pretty big leap forward as they’ve packed it with a total of eight models: six standard “dense” models and two massive MoE (Mixture-of-Experts) models. The idea here is to give developers a flexible toolkit they can use across everything from mobile apps and smart glasses to self-driving cars and robots.
Utilizing the latest ‘hybrid reasoning’ models that combine “thinking” mode for deep, multi-step tasks like solving math problems or writing code, and a “non-thinking” mode for quick, everyday responses, Alibaba wants its LLMs be autonomous and smart enough to know when to full-brainpower and when to keep it light and fast.
If you’re a developer using Qwen3 via API, you can even control how long it stays in thinking mode which is up to 38,000 tokens. That gives you the power to fine-tune performance versus speed. And if your use case is suitable for their largest MoE model (Qwen3-235B-A22B), it offers serious cost savings compared to other top-tier models, which is a big deal for anyone worried about cloud expenses.
Under the hood, Qwen3 was trained on a massive dataset of 36 trillion tokens, which is double what the previous version used. That’s translated into big improvements in areas like reasoning, tool use, and understanding multiple languages. It now supports 119 languages and does really well in both translation and following instructions in different languages.
It’s also built to work well with AI agents and tools like function-calling and following complex prompts are native features now. And if you’re into benchmarks, Qwen3 shines on the usual suspects: AIME25 for math, LiveCodeBench for coding, BFCL for tool use, and Arena-Hard for instruction-following.
To train these hybrid models, Alibaba went through a four-stage process, including chain-of-thought pretraining and reinforcement learning geared toward reasoning. As usual, Qwen3 is fully open-sourced and available on Hugging Face, GitHub, and ModelScope, with a demo up at chat.qwen.ai and integration into Model Studio coming in very soon.
