QwenLM/Qwen3.8-Flash-Next

阿里通义千问团队发布 Qwen3.8-Flash-Next：125B 参数 MoE（每 token 仅激活 6B），同时也是 Qwen4 架构的早期预览。训练成本降到 Qwen3.7-Plus 的 1/9，却在编程与办公场景上更强、更稳、更快。 技术报告： 博客： HF：

Excited to have contributed to building the next generation of Qwen models! Through architectural innovations, careful ablations, and extensive scaling experiments, we worked to make pretraining more…

Yeah it's strictly less expressive and not related to inference caching. Believe it's a training optimization in which they cache the smaller, down projected intermediate c instead of the query q in…

Confirmed: Qwen has released Qwen3.8-Flash-Next, an open-weight preview of the architecture behind Qwen4. It has 176B total parameters but activates only 6B per token. Qwen says it beats Qwen3.8-27B…

Alibaba open-sourced Qwen3.8-Flash-Next — 125B MoE, only 6B active per token. 1M-token context, $0.16/M input pricing, beats Opus 4.6 on 8/14 benchmarks. Alibaba itself flags which wins are…