Qwen3.6-35B-A3B

Qwen3.6-35B-A3B, published 15 April 2026, is the first open-weights release of Alibaba's Qwen3.6 line and an unusually honest one: on most of the vendor's own benchmark tables it does not beat Qwen3.5-27B, the dense model from the previous generation released seven weeks earlier. The comparison is worth reading carefully, because it says something about what a version bump means here. Qwen3.5-27B scores higher on SWE-bench Verified (75.0 vs 73.4), SWE-bench Multilingual (69.3 vs 67.2), SWE-bench Pro (51.2 vs 49.5), MMLU-Pro (86.1 vs 85.2) and Humanity's Last Exam (24.3 vs 21.4). Where the newer model pulls clearly ahead is in the sustained, tool-driven work an autonomous agent actually does: Terminal-Bench 2.0 51.5 against 41.6, QwenWebBench 1397 against 1068, NL2Repo 29.4 against 27.3. Against its direct predecessor at the same size and sparsity, Qwen3.5-35B-A3B, the gains are unambiguous — Terminal-Bench 40.5 to 51.5, QwenWebBench 978 to 1397, NL2Repo 20.5 to 29.4. The economics explain the design. This is a mixture of experts: 35.95 billion parameters in store, 256 experts, of which 8 routed plus 1 shared fire on any given token, so roughly 3 billion are active. It costs about an eighth of a dense 27B model to run per token while carrying more knowledge, which is the trade an agent making hundreds of tool calls per task wants. The layout is 40 layers as ten repetitions of three Gated DeltaNet blocks followed by one gated-attention block, hidden dimension 2,048, vocabulary 248,320, multi-token prediction, and a context window of 262,144 tokens natively that Alibaba says extends to 1,010,000. One detail the release notes do not mention: the configuration file still identifies the architecture as `qwen3_5_moe`. The version number moved from 3.5 to 3.6; the construction did not. Alibaba describes the release as built on community feedback and prioritising "stability and real-world utility", and the numbers bear that out — this is a post-training and reinforcement-learning revision, not a new model family. The visible product change is thinking preservation: reasoning context can now be retained from earlier messages in a conversation instead of being regenerated each turn. Vision comes as standard, as everywhere in this generation, with results essentially level against Qwen3.5-27B and modest gains in object grounding (ODInW13 50.8 against 42.6) and video understanding (VideoMMMU 83.7 against 82.3). Weights are Apache 2.0. All figures above are vendor-reported and have not been independently reproduced.

▸News

The newer Qwen loses to the older one in most tests — and the config file says why

9/2/2026 · Alibaba, karta modelu Qwen3.6-35B-A3B (Hugging Face)

▸Videos

No videos yet.

▸Reviews