Thread of 2 posts

jump to replies

I used #OpenCode with qwen/qwen3.6-35b-a3b to take care of this. All test cases passed and I verified all the different fixes. The tools are so much better nowadays. It’s like tools ⚒️ for tools 🛠️.

2 replies

back to top
Joe Cooper 🇺🇦 🍉 , @swelljoe@mas.to
(open profile)

@bryan The tiny models are getting so good. Qwen 3.6 is shockingly good for its size, same with Gemma 4. I'm astonished at what can run on a local computer and actually write code, use tools, etc. Self-hosted is approaching the frontier of a year ago, which is an amazing amount of progress. For me, Opus 4.5 is when it got Good Enough...so, I guess we're 6-12 months away from not actually needing a subscription (but still needing a big PC), assuming the leading devs keep releasing new models.

@swelljoe
Local models have really become usable recently. Not just tolerable but actually good. I was impressed with Gemma 4 as well, but I’ve noticed better results with Qwen3.6 (the majority of my recent code has been Swift).

I’ve had instances where Gemma 4 will get stuck in a loop bouncing between edits of Views in SwiftUI. Code change won’t compile, edit again trying a different strategy, still won’t compile, edits to the previous strategy that failed the first time, and repeat. I have to stop the loop, ask Opus to fix the mess, then I can usually continue.

Qwen seems much more capable for my needs right now. Thinking is shorter, and solutions work the first time. I recommend giving it a try. qwen/qwen3.6-27b almost similar if 35b is too big.