Qwen3.5-35B-A3B

Qwen3.5-35B-A3B, published on 24 February 2026, is the model that makes Alibaba's sparse approach easy to grasp. The weight files hold 35.95 billion parameters, but each token is routed to eight of 256 experts plus one shared expert, so roughly three billion parameters do the work at any moment. The cost of running it is set by that active figure; the memory it needs is set by the total. That trade lands well. On Alibaba's tables the model reaches MMLU-Pro 85.3, GPQA Diamond 84.2, HMMT February 2025 89.0 and SWE-bench Verified 69.2 — above GPT-OSS-120B, OpenAI's open model with three and a half times the parameters, on the first three, and close to the hosted GPT-5-mini on all of them. On TAU2-Bench, which measures whether an agent completes a multi-step customer task, it posts 81.2, the highest figure in the vendor's own comparison — ahead of its larger sibling Qwen3.5-122B-A10B at 79.5 and of the dense 27B at 79.0. Like everything in this generation it sees. Vision is trained in from the start rather than added as a separate line, and the model scores MMMU 81.4 and MathVision 83.9, the latter well clear of the previous generation's dedicated Qwen3-VL-235B-A22B at 74.6 — a vision specialist more than six times its size. The layout is 40 layers arranged as ten repetitions of three Gated DeltaNet blocks and one gated-attention block, each followed by the expert mixture; hidden dimension 2,048, expert width 512, vocabulary 248,320, trained with multi-token prediction. Unlike the smaller dense models of this generation, the output embedding is not tied to the input side. Context is 262,144 tokens natively, which the vendor says extends to 1,010,000. The weights are Apache 2.0, with the unmodified licence text in the repository, and the model is pulled from Hugging Face about 2.4 million times in a 30-day window — with a further 13.5 million on the FP8 build of its 3.6 successor, the most-downloaded checkpoint the Qwen team currently ships.

▸Videos

No videos yet.

▸Reviews