Qwen 3.8 Closes the Reasoning Gap, but Agentic Coding Remains a US Stronghold

The release of Qwen3.8-2.4T-A95B shifts the open-weight frontier—just not in the direction the headline scores suggest. Alibaba has shipped a 2.4 trillion-parameter sparse Mixture-of-Experts model with 95 billion active parameters per token, under a custom license that locks commercial scaling behind a $50 million revenue threshold. The weights arrived on Hugging Face around August 12–13, and the market responded immediately: Alibaba shares rose 7% in Hong Kong and 4.5% on the NYSE after the model’s early-August API launch.

Before accepting those numbers as proof of arrival, though, it is worth asking what the verified benchmarks actually confirm—and where the gap persists.

What the Verified Data Shows

On HuggingFace’s independently evaluated leaderboards, Qwen 3.8 posts 92.6 on GPQA Diamond and 86.6 on Terminal Bench 2.1. These are genuine general reasoning scores—competitive with GPT-5.6 Sol (94.1 and 88.8, respectively) and Opus 4.8 (92.0 and 84.6). The model has straightforwardly closed the reasoning gap with US frontier labs on standardized, cross-metric evaluation.

But beware the vendor-reported numbers dominating the coverage. Alibaba claims a 93.0 on PaperBench—higher than GPT-5.6 Sol’s 90.5, Opus 4.8’s 80.3, or Fable 5’s 88.8. That figure comes from Qwen’s own evaluation harness, with timeout and token-limit configurations that differ from those used for the US models. It is directionally interesting. It is not the same class of evidence as a HuggingFace leaderboard result.