Qwen3 30B A3B

Qwen3 30B A3B is a Mixture-of-Experts language model with an architecture built around selective expert activation. With 128 total experts in its pool but only 8 active during any forward pass, the model routes tokens through specialized sub-networks rather than engaging its full 30.5 billion parameters at once, making inference dramatically more efficient than a dense model of comparable size. The design stacks 48 transformer layers on a grouped-query attention mechanism, balancing computational throughput against the rich, layered representations needed for complex reasoning. A defining capability is its dual-mode operation: seamless switching between thinking mode for extended multi-step problem solving and non-thinking mode for efficient, general-purpose dialogue, all within a single model checkpoint.

The model progresses through both pretraining and post-training stages, with the later 2507 revision bringing measurable gains in instruction following, logical reasoning, text comprehension, mathematics, science, coding, and tool integration. Long-context understanding received particular attention, with native support extending to 256K tokens in the updated version. Over 100 languages and dialects are supported with strong multilingual instruction-following and translation capabilities. Human preference alignment was prioritized during refinement, yielding stronger performance in creative writing, role-playing, multi-turn dialogue, and open-ended subjective tasks. Agent capabilities were cultivated to enable precise external tool integration in both operational modes, achieving competitive results on complex agent-based benchmarks. At its scale, the design reaches state-of-the-art performance levels, matching the inference capability of larger models like QwQ-32B while its general capabilities substantially exceed earlier dense models of similar parameter count.