Qwen3 30B A3B

Qwen3 30B A3B is a causal language model built on a mixture-of-experts architecture, featuring 30.5 billion total parameters with 3.3 billion parameters activated during inference. Designed to balance depth and efficiency, the model supports a unique operational design that allows users to switch between a specialized thinking mode for complex tasks like mathematics and coding and a non-thinking mode for standard conversational interactions. This dual-mode capability is supported by a 48-layer structure and grouped-query attention, enabling the model to maintain high performance across diverse scenarios ranging from creative writing to intricate logical problem-solving.

The model benefits from extensive pre-training and post-training stages, resulting in significant advancements in instruction following and agent-based task execution. By integrating seamlessly with external tools, it excels in complex, multi-turn environments where precise reasoning is required. Its training lineage emphasizes human preference alignment, ensuring that the model remains engaging and natural in role-playing and dialogue. With support for over 100 languages and dialects, this model is positioned as a robust solution for developers seeking a balance between high-level reasoning performance and the operational efficiency of a smaller activated parameter count.