概念
也叫：MoE
一种稀疏激活的神经网络架构，通过条件计算降低推理成本同时保持模型容量。
最近变化
2026-06-04 · Nemotron 3 Ultra采用混合Transformer-Mamba MoE架构，550B总参仅激活55B，显著降低Agent任务计算开销。
Mixture-of-Experts (MoE) 被反复提及时，通常意味着它正在影响产品路线、开发者工作流或 AI 产业判断。这个页面把分散材料合并成一个可持续更新的观察入口。
已收录 4 篇与「Mixture-of-Experts (MoE)」相关的 AI 资讯和分析。
JetBrains 发布 12B MoE 架构模型 Mellum2，每 token 仅激活 2.5B 参数，推理速度超同类模型 2 倍以上，专为代码与文本任务优化，支持私有部署和 RAG 等高频低延迟场景。
入选理由：Mellum2 是 12B 参数 MoE 模型，每 token 仅激活 2.5B 参数，推理效率提升 2x+，适合高吞吐生产环境。
EMO是一种通过端到端预训练实现模块化涌现的混合专家模型，仅需12.5%的专家即可保持接近全模型性能。
入选理由：EMO 使用14B总参数、1B活跃参数，仅激活1/8专家即达近全模型性能。
This article highlights the advancements in small language models, specifically those with under 7 billion parameters, which can now run on consumer GPUs or even laptops. It emphasizes that these models are now capable of performing tasks that were previously only achievable by much larger models, thanks to improvements in training data quality, distillation techniques, and architectural innovations like Mixture-of-Experts (MoE). The article provides a curated list of the best small language models available on Hugging Face, along with their capabilities and benchmark scores.
入选理由：Small language models under 7 billion parameters are now capable of performing complex tasks previously reserved for much larger models.
NVIDIA Nemotron 3 Ultra已在Amazon SageMaker JumpStart上线，支持一键部署。该550B参数MoE模型专为长程Agent设计，推理速度提升5倍，成本降低30%，支持1M上下文。
入选理由：Nemotron 3 Ultra采用混合Transformer-Mamba MoE架构，550B总参仅激活55B，显著降低Agent任务计算开销。
与「Mixture-of-Experts (MoE)」经常一起出现的 AI 术语。
💡 想追踪「Mixture-of-Experts (MoE)」的长期趋势？去 实体雷达 · Mixture-of-Experts (MoE) 查看详细分析和跨材料问答。
