← sources

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

source: reddit (reddit:LocalLLaMA)published 2026-09-27 09:20ingested 2026-09-28 02:41

retained text

I run my own LLM while also having a Openai subscription. Also tried DeepSeek (latest flash now). I run [Qwen 3.8 flash Next](https://huggingface.co/albucino/Qwen3.8-Flash-Next-W4A16-FP8PLE) 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](http://pi.dev) and opencode until now) and the results amazed me! Suddenly AA benchmark made sense!

First test: The 3D Mario prompt test in Codex Cli blew me away. The best until now!

But real work is where you can see the difference! I took same project that Luna was working for days , and give it to both in paralel! And Qwen 3.8 Flash Next ran circle around luna. Previously, it failed to deliver results, with Qwen and DeepSeek as well in this project.

Now, I could say it's my go-to model!

P.S. For weebsearch, I've asked to port the [pi-smart-web-search](https://pi.dev/packages/pi-smart-web-search) to codex as a skill. It works amazing! (I should put it on git later). 



Maybe, I was using [pi.dev](http://pi.dev) wrong. Maybe there is an extensions that brings the same quality to it as Codex Cli. Does anyone know? 

document fetched 2026-09-28 00:39 · canonical https://reddit.com/r/LocalLLaMA/comments/1wrfp50/another_harness_matters_post_codex_cli_pi_and