Qwen 3.8 27B: a strong new model that overthinks

Alibaba’s Qwen research lab released Qwen 3.8 27B on Friday, an Apache 2 licensed vision-capable large language model with 27 billion parameters, according to Simon Willison’s blog. Willison writes that he had been anticipating the release, noting that 27B is a favorable size for running on a reasonably specced laptop and that the model’s predecessor, Qwen 3.6 27B, had impressed him when covered in April.

Willison links to Qwen’s published benchmark results on Hugging Face as part of his assessment. His post’s title states plainly that the model is excellent but defaults to what he describes as wildly overthinking things, a characterization that forms the center of his write-up, per the material provided.

The available summary text is truncated before Willison’s fuller discussion of the benchmarks and the overthinking behavior itself; the source material breaks off mid-sentence after referencing the benchmark-results section of the Hugging Face model page. Willison’s post is dated August 16, 2026, and appears on his personal blog, simonwillison.net, where he regularly tracks new open-weight model releases and their practical performance on consumer hardware.

No further quoted detail on the specific nature or extent of the model’s reasoning excess is available from the supplied excerpt.