Qwen3.5-27B

Qwen3.5-27B, published on 24 February 2026, opened Alibaba's Qwen3.5 generation and remains the most capable open-weights model the company has released at a size that fits on a single high-end accelerator. Its context in the catalogue is unusual. The previous Alibaba flagship with open weights, Qwen3-235B-A22B, held 235 billion parameters in store. On the vendor's own table the 27B dense model beats it on MMLU-Pro (86.1 vs 84.4), GPQA Diamond (85.5 vs 81.1), Humanity's Last Exam with chain of thought (24.3 vs 18.2), IFEval (95.0 vs 87.8) and IFBench (76.5 vs 51.7) — an 8.7-fold reduction in stored parameters with scores that went up rather than down. Against OpenAI's GPT-5-mini it leads on instruction following, long-document reasoning and agentic tool use (HLE with tools 48.5 vs 35.8), and against GPT-OSS-120B it leads almost everywhere. The one column where it clearly loses is competitive programming: 1,899 Elo on CodeForces against 2,160 for GPT-5-mini, 2,157 for GPT-OSS-120B and 2,146 for the older 235B model. Nothing else in the table drops that far behind, which makes it the honest limit of this model — algorithmic contest problems are still a capacity game. The architecture is dense: 27.78 billion parameters, 64 layers laid out as sixteen repetitions of three Gated DeltaNet blocks followed by one gated-attention block, hidden dimension 5,120, feed-forward intermediate dimension 17,408, vocabulary 248,320 tokens, multi-token prediction. Context is 262,144 tokens natively, extensible to 1,010,000 according to Alibaba. Vision is native to the generation rather than a separate line, and the model takes images and video alongside text. One discrepancy is worth recording. This model's own card reports 72.4 on SWE-bench Verified; the comparison table published with the newer Qwen3.6-35B-A3B in April gives the same model 75.0. Alibaba does not explain the difference, and the repository was updated the same day the newer card appeared — a re-run rather than a correction, most likely, but the reader should know that both figures come from the vendor and disagree by 2.6 points. Weights are Apache 2.0 with no added restrictions.

▸Videos

No videos yet.

▸Reviews