<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>GEEKR.Lee</title><description>Thinking &amp; Walking &amp; Writing about Technology &amp; Life</description><link>https://geekrlee-blog.pages.dev/</link><item><title>Pi：一个最小 agent harness，和我怎么把它跑在本地</title><link>https://geekrlee-blog.pages.dev/posts/2026-10/pi-coding-agent-notes/</link><guid isPermaLink="true">https://geekrlee-blog.pages.dev/posts/2026-10/pi-coding-agent-notes/</guid><description>Pi 1.2.0 的设计理念和架构：agent loop、JSONL 会话树、context 组装。记录本地部署、接 llama.cpp/Ollama 本地模型，以及 pi-web-ui 这个浏览器界面日常用下来的感受。</description><pubDate>Thu, 22 Oct 2026 05:00:00 GMT</pubDate></item><item><title>在四块 2080 Ti 上跑 Qwen3.8-Flash-Next（Strata）</title><link>https://geekrlee-blog.pages.dev/posts/2026-10/qwen3-flash-next-on-strata/</link><guid isPermaLink="true">https://geekrlee-blog.pages.dev/posts/2026-10/qwen3-flash-next-on-strata/</guid><description>4×RTX 2080 Ti + 188 GB RAM，用 Strata 跑 125B MoE：71–85 tok/s 解码、1.4K tok/s prefill，专家缓存驻留显存，HumanEval 40/41。附部署过程、coding bench、调参结论和踩坑清单。</description><pubDate>Thu, 22 Oct 2026 00:00:00 GMT</pubDate></item></channel></rss>