Last week I started a new series of videos: local AI models, tested on my computer, at the actual speed, when possible. The goal is to give my subscribers a very good understanding of what it is to run local, private AI on an entry level computer.

The first one is a heavily quantized version of Qwen 3.8 27B. It’s called Ternary Bonsai 2, and it is roughly 7.5 GB on disk. The normal Qwen 3.8 27B is around 50GB so it won’t fit on my 16GB M1 MacBookPro. Not even close. If you watched my series about how to choose an open weights model, then you know that size of the model must be smaller than computer RAM. I explain all this in episode 4.

Without further ado, here’s the video:

PrismML released Ternary Bonsai 2 ten hours before I recorded. I’ve been very enthusiastic at first, but after a couple more days of playing with it, I hit some limits, and realized that the team claims that the model retains 98% of the original Qwen performance must be taken with a grain of salt. I gave it a Three.js generation task and it worked for about 3 hours without producing relevant results.

Nevertheless, the model is still impressive. Especially when you think we couldn’t even dream about this level of performance a year ago.

Local, sovereign AI is evolving very fast. Once again, the video above is recorded on a 16GB M1 MacBookPro, not on a last generation Nvidia Blackbird, nor on a big memory Mac Studio. It’s just a tiny laptop.

I intend to make these series weekly, so if you have any suggestions about which models should I test next, drop a comment on that video. Of course, if you like this kind of content: like, share and subscribe!

Thanks for watching and spreading the word, you’re helping local, sovereign AI to come a little bit closer with every share.