Qwen3.8-Flash-Next released, with SGLang support on day one
Big thanks to @sgl_project for the day-0 support! 🙌 Qwen3.8-Flash-Next is ready to deploy with SGLa...
AI BriefingQwen has released Qwen3.8-Flash-Next, with deployment support from SGLang available on launch day. The model has a 125B main model, including 51B N-gram embeddings, and activates 6B parameters per token. The N-gram embeddings expand capacity with almost no additional computation and can reside in system memory using asynchronous prefetching. Its GDN+QSA hybrid attention balances efficiency on long tasks with retrieval accuracy.
Reported by 5 additional sources
- Qwen3.8-Flash-Next released: Qwen open-sources a multimodal MoE model, previewing the Qwen4 architectureSimon Willison’s Blog · August 27
- Alibaba’s Qwen team releases Qwen3.8-Flash-Next: a 125B multimodal MoE model with only 6B active parameters, previewing the Qwen4 architectureMarkTechPost · August 26
- Alibaba releases Qwen3.8-Flash-Next, focused on “extreme cost efficiency”The Decoder: AI News · August 26
- Qwen open-sources Qwen3.8-Flash-Next, with Day-0 support from SGLangLMSYS Blog · August 26
- Qwen3.8-Flash-Next goes open source: an early preview of the Qwen4 architectureQwen: Blog Retrieval · August 26
Source: Tongyi Qianwen / Qwen · x.com
Event follow-up1
View all event coverage- Qwen3.8-Flash-Next goes open source: an early preview of the Qwen4 architectureQwen: Blog Retrieval · August 26