The Current

Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index

A relatively small model matched larger competitors on a third-party benchmark, according to Simon Willison.

useful models · for technical · August 18, 2026

Writing on his blog on 17th August 2026, Simon Willison reported that Qwen 3.8 27B scored 52 on the Artificial Analysis Intelligence Index. According to Willison, that is the same score as GPT-5.6 Luna (max), and one point behind GLM-5.2 (max) and DeepSeek V4 Pro 0813 (max). Willison noted a contrast in model sizes: he wrote that the GLM model is 753B parameters and referenced DeepSeek at 1.6B (as stated in the source), while the size of Luna is unknown but 'presumably a whole lot bigger than 27B.' By comparison, Qwen 3.8 has 27B parameters. Willison described Qwen 3.8 27B as 'a truly astonishing model.' The post is a link post pointing to the underlying benchmark result. The comparison rests on a single independent source, Willison's write-up, rather than a vendor announcement or an aggregation of multiple outlets. The figures cited are drawn directly from that post.

  • Qwen 3.8 27B scored 52 on the Artificial Analysis Intelligence Index
  • That score matches GPT-5.6 Luna (max) and is one point behind GLM-5.2 (max) and DeepSeek V4 Pro 0813 (max)
  • Qwen 3.8 27B has 27B parameters, far fewer than the 753B GLM model cited
  • Reported by Simon Willison on 17th August 2026

What it means for you

A smaller AI model (27 billion parameters) reportedly scored as well as much larger ones on one benchmark. In plain terms: smaller models are catching up, which means cheaper-to-run AI that you could potentially host yourself may become good enough for real work. But a single benchmark score is not proof that it performs well on your specific tasks.

Try this

If you already run models locally or manage your own infrastructure, download Qwen 3.8 27B and test it against your actual use case rather than trusting the score.

Who should care

Developers and technical teams who self-host models or care about running AI cheaply on modest hardware.

Skip this if

You use hosted AI tools like ChatGPT or Claude and have no plans to run your own models — benchmark rankings between model versions won't change how you work.

Sources: Simon Willison — read the original

← All stories