Qwen3 30B A3B (Non-reasoning)

Qwen3 30B A3B (Non-reasoning) is a mid-sized language model developed by Alibaba Cloud's Qwen team, released as part of the Qwen3 family. This model utilizes a Mixture of Experts (MoE) architecture, featuring 30.5 billion total parameters with approximately 3.3 billion activated per token. It is specifically optimized for efficient, general-purpose dialogue and instruction following, serving as the standard "non-thinking" version of the architecture compared to its reasoning-enhanced counterparts.

The model was pre-trained on a massive corpus of 36 trillion tokens covering 119 languages and dialects. It supports a native context length of 128,000 tokens, which can be extended to 1 million tokens using specialized configurations. Its core capabilities include high-performance multilingual conversation, advanced code generation, mathematical problem solving, and complex tool-calling tasks within agentic workflows.

Designed for low-latency applications, this version prioritizes rapid response times and instruction alignment. It incorporates architectural refinements such as qk layernorm and global-batch load balancing for the MoE routing, ensuring stable performance across a wide range of tasks. The model is licensed under the Apache 2.0 license, facilitating both research and commercial use cases.