I’ve set up a local model to use with Codex. It runs at about the same speed as Claude with Sonnet at minimum effort.
It definitely does some dumb stuff, but so far I’ve been able to fix all of it with better instructions. We’ll see how it goes.
Why did I choose gpt-oss-20b?
I had Claude benchmark around 20 models on my machine and give me a report comparing execution time and quality.
After 7 hours of testing and reasoning, it concluded that gpt-oss-20b was the best fit for my setup.
I’m going to stick with it for now and see how it performs in actual day-to-day work.
Top comments (1)
Love this — running a local model day-to-day is the real test, benchmarks only tell you so much :)
Couple of questions: what's your setup (machine/specs)? Curious what it takes to run gpt-oss-20b at that speed.
And beyond the free tokens, what's been the biggest win so far — privacy, no rate limits, cost? Trying to work out if it's worth setting up on my end.