Qwen 3 30B A3B

MoE branch of Qwen3

This model moves Qwen3 into a sparse MoE serving geometry while keeping the same user-facing thinking/non-thinking framing.

Low active path

Only 3.3B parameters are activated per token out of 30.5B total, which is the central deployment distinction versus the dense Qwen3 line.

Agent and reasoning focus

Qwen still positions the model for reasoning, instruction following, and complex agent workflows rather than only general chat.

MoE family branch

Qwen3 includes dedicated MoE models alongside the dense line, keeping the same user-facing thinking/non-thinking framing while changing the serving geometry materially.

Sparse activation

The MoE releases expose total and activated parameter counts separately, which is the key deployment distinction versus the dense Qwen3 models.

Long-context packaging

The base MoE releases are published with 32K native context and 131K support with YaRN, while the 2507 update is packaged at 256K native context.

Reasoning with lower active compute

The MoE line is for users who want larger total capacity without paying dense-model active compute per token.

Agent and tool use

Qwen still positions the MoE branch around agent workflows, tool calling, and mixed reasoning/general dialogue use.

Large multilingual serving

Useful when you want very large-capacity multilingual serving without moving to a purely dense 70B+ model.