Alibaba Open Sources Qwen3.8-Flash-Next with DeepSeek-Inspired Architecture
量子位 · wechat · 2026-08-27
Alibaba released the weights for Qwen3.8-Flash-Next, a preview of the Qwen4 architecture. The 125B MoE model features 51B N-gram Embedding parameters (inspired by DeepSeek's Engram), significantly reducing training costs to 1/9th while improving performance. It also upgrades attention (QSA), residual connections, and the optimizer (Muon). The model supports 1M token context, offers API costs roughly 1/12th of the Max version, and is compatible with OpenAI/Anthropic protocols.
More from Models
- Zhipu GLM-5.3 open weights releasing in 22 hours — Yuchenj_UW · 2026-08-27
- AI models show more creativity when talking to each other than in assistant persona — nabeelqu · 2026-08-27
- OpenRouter leaderboard: Real token consumption data outweighs media hype — sujingshen · 2026-08-27
- Qwen 3.8-Next Released with Detailed Technical Report on Architecture — nrehiew_ · 2026-08-27
- OpenAI's token efficiency may stem from training budget awareness — teortaxesTex · 2026-08-27
- Qwen 3.8 UX improvement: Simple operations no longer cause token redundancy anxiety — infieldmitt · 2026-08-27