I can't believe it, but I'm able to do my daily software development work on this model. We have a 500-700k line of code enterprise software suite that I'm devving for 60 hours a week.
I've been hunting for a cursor replacement for a little bit now, and was previously toying with Kimi 2.6 and deepseek 4 pro and flash.
There are some minor issues I've had with each of those, and Q3.6:35b-a3b actually feels the best for me, anecdotally, of all of them.
I can't articulate how insanely excited and shocked I am.
I've been hearing the hype here for a bit and I have to say it lived up to it. I could run this model locally, but I don't have the hardware for it, so for now I'm using it on openrouter at ~$0.08/1M tokens averaged out for our usage (what we're actually getting billed after caching and whatever is figured out).
That's so insanely cheap for a model that can actually understand what I need it to with this workload / use case, and can accept image input / screenshots.
If you haven't tried this model, I implore you, take a look at it. It's shockingly good.
The only thing that I miss from Cursor at this point is the cloud agents functionality, and the high throughput they have on auto/Composer 2.
[link] [comments]