Qwen 3.8 Flash-Next: 6B-Active Open Model Beats Claude Opus 4.6 Max

kimmonismus · x · 2026-08-26

Alibaba's Qwen team released Qwen3.8-Flash-Next, an extremely sparse MoE model. While it has a total of 125B parameters (plus 51B n-gram embeddings), only 6B parameters are active per token. The model performs excellently on multiple benchmarks, beating Claude Opus 4.6 Max on 8 comparable metrics, including SWE-bench Pro (62.5), SWE-bench Multilingual, and CoWorkBench. It also supports native 256K context extendable to 1M, with significantly faster inference speeds at long context (7.6× faster prefill, 4.9× faster decode).

Related event: Alibaba Qwen releases Qwen3.8-Flash-Next, a 125B sparse MoE previewing Qwen4 architecture(8 posts)→

Original post →

More from Models

Models channel →