This is episode 13 of the series in which I test local models that actually fit on a 16 GB MacBook Pro. Today’s model is Nanbeige 4.2, the 3B version.

It comes from a company connected to Boss Zhipin, the largest Chinese online recruiting platform. That makes it a bit unusual, because most of the models I’ve tested so far come from the big labs. This one was trained with their own data and their own recipes, so the output might not feel exactly like another Qwen or DeepSeek variant.

The 4Bit Quant Was Good, But Slow

I ran the smaller 3B base model as a 4-bit abliterated quant. The file is only 2.1 GB. On a 16 GB Mac that leaves a lot of free memory for other apps. The trade off is obviously speed. This was the slowest model I’ve tested in the series. The thinking phase is long and visible, which is interesting to watch, but it also means a simple text question took about two minutes end to end. A small coding task (write a utility that counts words in files in a folder) took roughly ten minutes of thinking before it finished with a working script.

The results themselves were solid — above average for the size — and the model did produce usable code. It just requires more time than anything else I’ve run so far.

Fair warning: I only tested the 4-bit version, I reckon higher quants might be faster at the cost of a few extra gigabytes of RAM, but I didn’t check that.

Nanbeige 4.2 3B is usable on a 16 GB Mac if you can accept the speed. It’s small, the output is decent, and the thinking process is worth watching once. Just don’t expect quick replies.

If you like this kind of content, like, share and subscribe to my channel!