PulseAugur
中
实时 03:30:11
English(EN) The great sparsification: 2026 open-weight MoE models now activate 1 to 9 percent of their weights

2026年MoE模型分化:总参数与激活参数创造双重预算

2026年,开放权重专家混合(MoE)模型在总参数量与激活参数量之间出现了显著分化。这意味着决定内存需求和加载时间的总权重远大于决定每个token计算成本的激活权重。例如,Kimi K3拥有2.8万亿总参数,但仅激活约1040亿;而DeepSeek的V4.1-Flash拥有5520亿总参数,但仅激活80亿。这一趋势使得模型能够存储海量知识,同时保持每个token的计算成本可控,但通过创建独立的内存和计算预算,使预算和部署变得复杂。 AI

影响 MoE模型的这种架构转变,由于内存和计算成本现在是解耦的,因此需要新的策略来预算和部署AI系统。

排序理由 讨论开放权重模型架构和性能指标的趋势。[lever_c_demoted from research: ic=1 ai=1.0]

在 dev.to — LLM tag 阅读 →

AI 生成摘要 · Google Gemini · 来自 1 个来源。 我们如何撰写摘要 →

2026年MoE模型分化:总参数与激活参数创造双重预算

本文如何被排名

Signal score
23 / 100
Composite score across the factors below. Higher = stronger signal that this story matters right now.
Newsworthiness bucket
Tool
讨论开放权重模型架构和性能指标的趋势。[lever_c_demoted from research: ic=1 ai=1.0]
Source corroboration
Single-source cluster
Only one publisher covered this so far. Single-source stories can still rank when the publisher is high-authority, but they lack cross-source corroboration.
Topics
model release, infra
Editorial topic classification. Feeds into how the story surfaces on /topic/<slug> hub pages and into the per-entity coverage mix.
AI-industry relevance
High
Clearly on-topic for AI-industry coverage.
Story freshness
Breaking (< 6h)
Fresh story with cross-source coverage still developing. Ranking may shift as more sources report.

完整方法见我们的编辑标准。

报道来源 [1]

  1. dev.to — LLM tag TIER_1 English(EN) · ai maya ·

    大稀疏化:2026年开放权重MoE模型现激活1%至9%的权重

    <h2> TL;DR </h2> <p>In 2026 the headline parameter count of an open-weight model stopped telling you how much compute it burns per token. The dominant design is a sparse Mixture of Experts (MoE) where the total weight budget keeps climbing into the trillions while the <em>active<…