#AGI 阿里巴巴發佈全新 Qwen3-Next 模型架構，以混合注意力機制與高稀疏專家模型（MoE）大幅提升效率。最新的 80B 參數模型僅啟動 3B 參數即可推理，不僅在長上下文任務上突破極限，更將訓練成本壓縮至不到 Qwen3-32B 的 10%。此外，Qwen3-Next 系列同時開源，並推出多語種高精度語音識別工具 Qwen3-ASR-Flash。
Why it matters: Qwen3-Next delivers top-tier performance at a fraction of the computational cost, making advanced AI models more accessible for research and deployment.
The big picture: Alibaba’s push into ultra-efficient architectures highlights a broader shift in AI toward scaling responsibly—balancing capability, cost, and sustainability in the race for AGI.
Full article https://www.alizila.com/qwen3-next-a-new-generation-of-ultra-efficient-model-architecture-unveiled/
📌 一杯咖啡價錢連接 Web3 世界 https://patreon.com/wanszezit
Category: 剪報