Qwen3.8-27B

The call

The August 2026 dense refresh, and the most-adopted open-weight model of the month by a wide margin — 267K downloads and 10.4k likes in its first twelve days, with 480 community quantizations already published. It supersedes Qwen 3.6-27B at the same size, the same Apache 2.0 licence and the same 24 GB target, but the architecture underneath is different: a hybrid layout of 16 × (3 × Gated DeltaNet → 1 × Gated Attention), the same linear-attention family Alibaba used in Qwen3.8-Max. Three quarters of its 64 layers use linear attention whose state does not grow with context, so the KV cache at long context is roughly half what the dense 3.6-27B needs. It is also a native vision-language model — images and video in, not a text model with an adapter bolted on — and it is trained with multi-token prediction, so speculative decoding works without a separate drafter.

When not to use: When you want the fastest tokens/sec on a 24 GB card — the 3.6-35B-A3B MoE still wins on raw decode speed because only 3B parameters activate per token. Also treat the benchmark tables on the model card with the usual caution: they are vendor-run, evaluated with Alibaba's own harness, and we have not seen independent numbers yet. The reason this displaces Qwen 3.6-27B in our picks is not the benchmark claims — it is that it is the same vendor's direct successor at the same size and licence, with an order of magnitude more adoption and a materially cheaper long-context profile.