Transitioning from proprietary to open source models and harness

Posted by nopickles_@reddit | LocalLLaMA | View on Reddit | 3 comments

Hey all, I’ve been using Claude Code with Opus and Sonnet but as you all know the rate limits as well model capabilities have degraded significantly. To that end I want to transition to the open source eco system but I’m very lost. Here are my questions I’m looking for help with:

Which open source models to use? I know GLM 5.1 that just dropped is on par with Opus 4.6, but what about a replacement for Sonnet for traditional coding and stuff? I’ve heard about Kimi and Minimax etc
Is OpenCode a better harness for the open sourced models? or should I stick with Claude Code?
Finally, is there like a centralized place I can check to track the new open source releases, scores, usages etc?

Thanks a lot in advance

[-]

ttkciar@reddit

I would posit that Open Code is the best of the open source agentic codegen applications, but Claude Code is the slightly better application.

If you're okay with running closed-source software, Claude Code should be fine, though I do not have a lot of experience with it compared to Open Code.

As for open-weight models which are a step down from GLM-5.1, you might look at GLM-4.5 or GLM-4.6, which avoid the guardrail weirdness ZAI introduced with GLM-4.7.

If you decide to try running a codegen model on your own hardware, you should start with GLM-4.5-Air, which is my go-to for local codegen inference. It falls far short of Sonnet, but in my experience it beats out other models in the 120B range (GPT-OSS-120B, Qwen3.5-122B-A10B, Devstral 2 Large 123B) especially in instruction-following competence.

Ryoonya@reddit

GLM 5.1 isn't good as Opus 4.6 on claude code.

But it is on par with Sonnet 4.6.

You will not find anything open source as 4.6 right now.

laterbreh@reddit

Heres the best advice: Try it. Go on openrouter and hook it up to opencode. Thats it. Thats the answer.