James O'Beirne publishes an opinionated guide to running SOTA LLMs locally

Category: Pulse
Status: discussion
Sources: local-llm guide, HN discussion
Summary: A guide titled "Everything I know about running LLMs locally" reached the front page (309 points). It lays out three hardware budget tiers (about 2k USD, 40k USD, and higher), recommends specific open-weight models such as Qwen and GLM per tier, and covers GPU optimization and speech-to-text setup, with the author noting nothing outside the tables was written by AI.
Comments: Commenters debated the price-to-capability tradeoffs, with several arguing an M5 MacBook Pro with 48GB shared memory competes with a 2x RTX 3090 build at similar cost, and others noting the near-Opus GLM 5.2 tier realistically needs far more than the stated 40k USD (closer to 8x H200) unless heavily pruned and quantized. Some flagged calling Qwen3.6-27B state-of-the-art as a stretch.