Qwen3.6-27B

Qwen3.6-27B, released on 22 April 2026, is the model that makes the case for dense architectures at a moment when almost every frontier release has gone sparse. It carries 27.78 billion parameters, all of them active on every token, and on Alibaba's own benchmark tables it beats Qwen3.5-397B-A17B — the company's previous-generation open-source flagship, fifteen times larger in total parameter count — across every major coding benchmark. The margins are not marginal. SWE-bench Verified 77.2 against 76.2, SWE-bench Pro 53.5 against 50.9, SWE-bench Multilingual 71.3 against 69.3, Terminal-Bench 2.0 59.3 against 52.5, NL2Repo 36.2 against 32.2. On SkillsBench the gap is a chasm: 48.2 against 30.0. On QwenWebBench, the vendor's internal front-end generation rating, 1487 against 1186. Two of those numbers are worth stating separately, because they are measured against Claude 4.5 Opus rather than against an open model: Terminal-Bench 2.0 is a tie at 59.3, and SkillsBench goes to the Chinese model, 48.2 against 45.3. What this does not mean is that a 27B model has caught up in general. Read the knowledge rows and the picture reverses: MMLU-Pro 86.2 against Opus's 89.5, Humanity's Last Exam 24.0 against 30.8, SuperGPQA 66.0 against 70.6. The model has been trained hard on the narrow, verifiable, tool-driven work that reinforcement learning can grade — patching a repository, driving a terminal — and that is where the parameter count stops mattering. Broad recall still costs parameters, and it shows. The construction is the same Gated DeltaNet layout as the rest of this generation: 64 layers arranged as 16 repetitions of three linear-attention blocks followed by one gated-attention block, hidden dimension 5,120, vocabulary 248,320, multi-token prediction, 262,144 tokens of context that Alibaba says extends to 1,010,000. Vision comes as standard. One detail carried over from the sibling release and still unmentioned in the notes: the configuration file identifies the architecture as `qwen3_5`, not 3.6 — the version number moved, the construction did not. Weights are Apache 2.0, downloadable from Hugging Face and ModelScope with an FP8 checkpoint alongside; resellers list the hosted model at roughly $0.60 per million input tokens and $3.60 per million output. All benchmark figures above are vendor-reported and have not been independently reproduced.

▸News

A 27B model beats Alibaba's own 397B flagship — and ties Claude 4.5 Opus on two agent tests

9/2/2026 · Alibaba, karta modelu Qwen3.6-27B (Hugging Face) i komunikat qwen.ai

▸Videos

No videos yet.

▸Reviews