<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
<channel>
  <title>海外ニュース速報 - ローカルAI</title>
  <link>https://648news.com/local_ai/</link>
  <description>海外ニュース速報</description>
  <language>ja</language>
  <lastBuildDate>Sat, 19 Sep 2026 13:29:36 +0000</lastBuildDate>
  <atom:link href="https://648news.com/local_ai/feed.xml" rel="self" type="application/rss+xml"/>
  <item>
    <title>RTX 3060 Ti 8GBで63モデル実測 帯域幅の壁とMoEオフロードの実像</title>
    <link>https://648news.com/local_ai/2026/09/19-rtx-3060-ti-8g-local-llm-benchmark</link>
    <guid isPermaLink="true">https://648news.com/local_ai/2026/09/19-rtx-3060-ti-8g-local-llm-benchmark</guid>
    <pubDate>Sat, 19 Sep 2026 06:00:00 +0900</pubDate>
    <category>ローカルAI</category>
    <description>GitHubユーザーBBQ2077が2026年9月18日に公開したベンチマークレポートは、RTX 3060 Ti 8GB搭載PCで63種類のローカルLLMを実測した結果をまとめたものだ。</description>
  </item>
  <item>
    <title>Accio-Labが長期エージェント向けコンパクトMOEモデルOccamy-1.0を公開</title>
    <link>https://648news.com/local_ai/2026/09/19-accio-lab-occamy-1-0-release</link>
    <guid isPermaLink="true">https://648news.com/local_ai/2026/09/19-accio-lab-occamy-1-0-release</guid>
    <pubDate>Sat, 19 Sep 2026 06:00:00 +0900</pubDate>
    <category>ローカルAI</category>
    <description>2026年9月18日、Accio-Labが長期ステートフルタスク特化のコンパクトMOEモデルOccamy-1.0をHugging Faceで公開した。総パラメータ35Bながらアクティブ3BのMixture-of-Expertsアーキテクチャを採用し、ローカル推論の実用性を追求している。</description>
  </item>
  <item>
    <title>Qwen3.8-Flash-Next EXL3をWindowsの消費向けGPU2枚で動作させる独自パッチとベンチマーク</title>
    <link>https://648news.com/local_ai/2026/09/19-qwen-flash-consumer-lab-windows-experiment</link>
    <guid isPermaLink="true">https://648news.com/local_ai/2026/09/19-qwen-flash-consumer-lab-windows-experiment</guid>
    <pubDate>Sat, 19 Sep 2026 06:00:00 +0900</pubDate>
    <category>ローカルAI</category>
    <description>jiafeimao1990は2026年9月18日、ExLlamaV3 v1.5.0をベースにした独自パッチを用い、RTX 5060 Tiを2枚搭載したWindows PCでQwen3.8-Flash-Next EXL3モデルを動作させる実験結果をGitHubに公開した。</description>
  </item>
  <item>
    <title>16GB GPUで27Bモデルの256Kコンテキストを実現するTierKVの公開</title>
    <link>https://648news.com/local_ai/2026/09/18-tierkv-llama-cpp-27b-256k</link>
    <guid isPermaLink="true">https://648news.com/local_ai/2026/09/18-tierkv-llama-cpp-27b-256k</guid>
    <pubDate>Fri, 18 Sep 2026 06:00:00 +0900</pubDate>
    <category>ローカルAI</category>
    <description>2026年9月17日、arczhiがllama.cppのフォーク「TierKV」を公開した。VRAM、システムRAM、SSDの3層でKVキャッシュを管理し、16GBのGPUで27Bモデルの256Kトークンコンテキストを維持しながら高速な推論を実現する。</description>
  </item>
</channel>
</rss>