Qwen3 Technical Report: A Unified MoE Large Model for Thinking and Non-Thinking Modes
In May 2025, Alibaba released the Qwen3 series of large language models, ranging from 0.6B to 235B parameters, including dense and MoE architectures. The core innovation is the unification of thinking mode (complex multi-step reasoning) and non-thinking mode (fast response), supporting dynamic switching and a thinking budget mechanism that adaptively allocates computational resources. The models achieve state-of-the-art results on benchmarks for code, mathematics, and agent tasks, with multilingual support expanded from 29 to 119 languages, all open-sourced under Apache 2.0.
Qwen3 is one of the most advanced open-source large models, for the first time unifying deep reasoning and fast response within a single model, allowing users to dynamically adjust reasoning depth based on task complexity without switching models. Its thinking budget mechanism enables flexible trade-offs between latency and performance, which is particularly critical for cost-sensitive enterprises. Multilingual expansion to 119 languages significantly enhances global usability. The MoE 235B model rivals or surpasses closed-source models on multiple benchmarks, marking a new height in open-source model capabilities.
Qwen3 adopts a hybrid architecture, with dense models (0.6B-72B) and MoE models (A14B-235B) sharing a unified training framework. The key innovation is a learnable gating mechanism for switching between thinking and non-thinking modes: in non-thinking mode, the model directly generates answers; in thinking mode, it first generates an internal reasoning chain before output. The thinking budget is implemented by controlling the number of reasoning steps or tokens, and users can specify the budget via a chat template. Training uses two stages: first pre-training on large-scale corpora, then reinforcement learning to optimize mode switching and budget usage. Evaluations show Qwen3-235B achieves 96.8% on MATH-500 and 92.3% on HumanEval, surpassing Qwen2.5-72B. Multilingual capability is achieved through cross-lingual transfer learning, maintaining high BLEU scores across 119 languages.
Qwen3's open-source strategy will accelerate global AI application deployment, especially for enterprises requiring multilingual support (e.g., cross-border e-commerce, international customer service). The unified thinking/non-thinking mode reduces model selection complexity, allowing enterprises to handle both simple Q&A and complex reasoning tasks with a single API. The inference efficiency advantage of the MoE architecture (reducing computational cost by 50% compared to dense models of similar scale) will drive more enterprises to adopt open-source models as alternatives to closed-source APIs.
It is recommended to compare Qwen3 with existing closed-source APIs on multilingual customer service, code, and complex reasoning tasks, making decisions based on accuracy, latency, and total cost of computing and operations. The thinking budget can be used for tiered services, but real traffic should be used to verify whether routing reduces the unit effective task cost.
Focus on the actual deployment effectiveness of Qwen3 in agent and tool-calling scenarios, especially the cost control capability of the thinking budget mechanism in long-chain reasoning tasks. Observe whether the community develops vertical domain fine-tuned models based on Qwen3, and the real performance of its multilingual capability on low-resource languages. Automated optimization of the thinking budget (e.g., dynamic adjustment based on task difficulty) will be a subsequent research hotspot.