Simon Willison reviews Qwen 3.8 27B, an Apache 2-licensed vision-capable LLM from Alibaba with self-reported benchmarks beating both its predecessor Qwen 3.6 27B and the closed-weight Qwen 3.7-Plus. However, he highlights that the model's default reasoning effort of 'xhigh' causes it to wildly overthink simple tasks, exhausting context limits quickly—suggesting users should lower the reasoning_effort for practical use.
Background
Qwen 3.8 27B is a mid-size open-weight vision-language model from Alibaba, following in a lineage of increasingly capable Qwen models. The 27B parameter size targets consumers who want strong model performance on local hardware like Mac M-series chips or NVIDIA DGX Spark.
- Source
- Lobsters
- Published
- Aug 17, 2026 at 10:39 PM
- Score
- 7.0 / 10