Decided to try the "svg of pelican riding a bike" test locally on my base M5 Macbook Pro. Here's the result made by Qwen 3.6 35b A3B running via oMLX with pi.
64tps avg decode
The trick is to leverage the model's native vision ability to let it actually see what it's working on
Looks like Bonsai 2 27B is all hype. It's the same chonic overthinker as base Qwen 3.8, but instead of iterating itself to passable results, it either fails completely or delivers something uninspired.
Fine Tunes of Qwen 3.6 MoE are still the way to go.
youtu.be/OA5cICIzD-c
Not sure if anyone mentioned this already, but MacOS 27 comes with a built in framework for local AI called fm
It even has a small 4B parameter model baked in that you can play around with
Been trying to use Qwen 3.8 27b lately. Unironically feels like an elaborate meme.
It’s better than something like Gemma4, but it’s not confident in anything it does. It’s just second guessing itself over and over and over like an abused animal.
Obviously it’s a step in the right direction for local LLMs, but the whole “Frontier level performance” comments from people is a straight up meme.
(And Yes. I’m using it with reasoning set to Medium)