Ask HN: Open-Source Coding Model and Harness at Claude Sonnet / Opus Level Perf?
3 comments
Re: open-source harnesses, ForgeCode appears to be pretty good (currently, #1 on Terminal Bench 2.0 -- https://www.tbench.ai/leaderboard/terminal-bench/2.0). Re: open models, Kimi K2.6 might be a good place to start, but admittedly I'm not too sure how it'd compare to Sonnet / Opus for your use case.
It's terrible, don't waste your time like I did. That bench means nothing by the way.
I have tried Gemma 4 on Ollama (Google's SOTA open source model that released Apr 2026), and it is absolute garbage. Stay away.
--can not agree more.
Some Context: - I have about 20GB-30GB of space left on my Macbook Pro, which I can use for downloading a local model. I do not want to use Ollama Cloud and pay yet another $20/mo subscription. - I have tried Gemma 4 on Ollama (Google's SOTA open source model that released Apr 2026), and it is absolute garbage. Stay away.