I got a 64gb memory mac about a month ago and I've been trying to find a model that is reasonably quick, decently good at coding, and doesn't overload my system. My test I've been running is having it create a doom style raycaster in html and js
I've been told qwen 3 coder next was the king, and while its good, the 4bit variant always put my system near the edge. Also I don't know if it was because it was the 4bit variant, but it always would miss tool uses and get stuck in a loop guessing the right params. In the doom test it would usually get it and make something decent, but not after getting stuck in a loop of bad tool calls for a while.
Qwen 3.5 (the near 30b moe variant) could never do it in my experience. It always got stuck on a thinking loop and then would become so unsure of itself it would just end up rewriting the same file over and over and never finish.
But gemma 4 just crushed it, making something working after only 3 prompts. It was very fast too. It also limited its thinking and didn't get too lost in details, it just did it. It's the first time I've ran a local model and been actually surprised that it worked great, without any weirdness.
It makes me excited about the future of local models, and I wouldn't be surprised if in 2-3 years we'll be able to use very capable local models that can compete with the sonnets of the world.
[link] [comments]