The July refresh of Alibaba's small MoE model: the hybrid thinking switch is gone, the context grows eightfold and the benchmark scores jump by tens of points.
Three months after the April release of Qwen3-30B-A3B, Alibaba published an updated model under a very similar name — and with a different design decision behind it. The original carried one set of weights that could be switched between a thinking mode and a direct mode. This edition drops the switch: it runs only in non-thinking mode and no longer emits reasoning blocks at all. A separate reasoning edition followed in August, so the family that started as one hybrid model ended the summer as two specialised ones.
The physical shape of the network is unchanged: 30.5 billion parameters, 128 experts across 48 layers, 8 experts routed per token for about 3.3 billion active. What changed is training and context. The window grows from 32,768 tokens natively to 262,144 — eight times more, without the YaRN workaround the April model needed for long inputs.
The maker's own comparison table shows how much the retraining bought, measured against the April model in non-thinking mode: AIME25 rises from 21.6 to 61.3, Arena-Hard v2 from 24.8 to 69.0, GPQA from 54.8 to 70.4 and MMLU-Pro from 69.1 to 78.4. Those are self-reported figures and the tests reward different things, but the direction is consistent across all of them. Weights remain Apache 2.0 with commercial use included; the repository records around 780,000 downloads in thirty days.