Qwen/Qwen3-30B-A3B-Instruct-2507

Qwen3-30B-A3B-Instruct-2507 is a mixture-of-experts language model built around 128 experts with 8 activated per token during inference, keeping active computation at 3.3 billion parameters while leveraging a much larger total capacity. The architecture features grouped query attention with 32 query heads and 4 key-value heads across 48 layers, all operating in a deliberate non-thinking mode that skips internal reasoning traces and produces direct responses. This design prioritizes efficient, high-quality instruction-following without the overhead of generating intermediate thinking steps.

The model went through both pretraining and post-training phases, with post-training bringing measurable gains in instruction following, logical reasoning, mathematical problem-solving, coding, and multilingual comprehension. Community quantizers have made the model available in compressed formats like AWQ INT4, bringing memory requirements down significantly while retaining benchmark performance near original perplexity levels. Its enhanced alignment with user preferences and 256K-context capability make it well-suited for complex, open-ended tasks where quality and helpfulness matter more than showing work. Security evaluators have assessed its robustness across dozens of vulnerability types, reflecting its growing adoption in production applications.