This is the second half of a decision Alibaba made in the summer of 2025. The April Qwen3-30B-A3B held a switch between thinking and direct answering in one set of weights; the July instruction edition kept only the direct mode, and this edition kept only the thinking mode. It cannot be talked out of reasoning — the chat template opens a thinking block automatically, which is why its output often shows a closing tag without a visible opening one.
The network is unchanged from April: 30.5 billion parameters, 128 experts across 48 layers, 8 routed per token for roughly 3.3 billion active. Context grows to 262,144 tokens natively. The maker warns plainly that this version thinks for longer than its predecessor and recommends it only for genuinely hard problems — a cost that shows up as tokens on the bill.
Against the April model in the same mode the maker reports AIME25 rising from 70.9 to 85.0, HMMT25 from 49.8 to 71.4, LiveCodeBench v6 from 57.4 to 66.0 and Arena-Hard v2 from 36.3 to 56.0. The number worth pausing on is the first one: 85.0 on AIME25 is above the 81.5 the same table gives the maker's own 235B-parameter flagship in thinking mode, and well above the 72.0 for Gemini 2.5 Flash Thinking. These are self-reported figures on a single maths competition set, and the same table has the model losing clearly to Gemini on GPQA (73.4 against 82.8) — but a 30B model beating a 235B one on any headline test is the kind of result that shaped how the open-weights market thought about model size that year. Weights are Apache 2.0 with commercial use included.