Qwen 3
Alibaba's next-generation open-weight LLM family
5.
Definition
Qwen 3 is Alibaba Cloud's next-generation family of open-weight large language models, introducing hybrid 'thinking' and 'non-thinking' reasoning modes within the same model along with mixture-of-experts variants, aimed at improved reasoning, efficiency, and multilingual coverage over Qwen 2.5.
Overview
Qwen 3 builds on the Qwen 2.5 lineage with a redesigned training pipeline and a notable architectural feature: a single model can switch between a fast 'non-thinking' response mode for simple queries and a slower 'thinking' mode that performs extended chain-of-thought reasoning for complex problems, controllable by the user or application rather than requiring a separate reasoning-specific model. This hybrid approach aims to give developers a single deployable model that can handle both latency-sensitive and reasoning-heavy workloads. The Qwen 3 family includes both dense models and mixture-of-experts (MoE) models at various sizes, with the largest MoE variants activating only a fraction of their total parameters per token to keep inference costs manageable while retaining large model capacity. Training data expanded further in scale and language coverage compared to Qwen 2.5, with Alibaba reporting substantial gains on reasoning, coding, and agentic tool-use benchmarks. Like its predecessor, Qwen 3 is released with open weights across most size tiers, supporting self-hosting, fine-tuning, and integration through common inference frameworks and platforms such as Hugging Face and Ollama. The combination of open licensing, competitive benchmark performance, and the thinking-mode toggle has made Qwen 3 a popular choice for developers building agentic applications that need to balance response speed against reasoning depth within a single deployed model.
Key Features
- Hybrid 'thinking' and 'non-thinking' modes controllable per request
- Both dense and mixture-of-experts (MoE) model variants
- Expanded multilingual training data over Qwen 2.5
- Improved reasoning, coding, and agentic tool-use benchmark performance
- Open-weight release across most size tiers
- Long context window support
- Native function/tool calling for agentic pipelines
- Broad availability across Hugging Face, Ollama, and cloud inference providers