2026-09-27 · America/Los_Angeles · 社区动态 · #6

Another "Harness matters" post (codex cli > pi and opencode)

I run my own LLM while also having a Openai subscription. Also tried DeepSeek (latest flash now). I run Qwen 3.8 flash Next at an amazing speed on my 2x3090 + Ram! But local LLM never did worked for me outside some demos like build me a "3D Mario Game, multistage" which I've been using to test LLM's for a long time. At serios work, they never even compared with GPT 5.2 or lately, 5.6 Luna, which is worse in the benchmarks. Until last night! I've asked gpt 5.6 luna to configure codex cli for local llm! (I've been using pi.dev and opencode until now) and the results amazed me! Suddenly AA…

热度 46.8 / 100；排名与评分保留该期记录。