For the 15 episode of my series in which I test local AI models that fit on a 16GB MacBook Pro, I chose Raon-1 nano from KRAFTON (the PUBG company). This is a Korean model, trained from scratch, not a distillation of a model by another provider. It’s a dense model, not a MoE, but it has an interesting architecture: out of its 3.1 billion parameters, only 1.5B are effective, thanks to per-layer embeddings (if you don’t understand what that means, it’s ok, just know it’s faster). The nano version I tested its distilled from their own 30B-A4B Raon-1 mini. In my tests I used the Q8_0 quant for a total 3.4 GB on disk.

The Fastest Model So Far

This is by far one of the fastest models in my series. The dolphin answer came back almost instantly (~12 tok/s). The coding task, the same word-count utility finished in about 30 seconds (~17 tok/s) with a working script and usage notes. Because the video was under 3 minutes, which YouTube treats as a short, I threw at it a third task: a simple SwiftUI iOS calorie-tracker skeleton. That took roughly 70 seconds and produced something I could compile with only minor fixes.

Quality is high enough for day-to-day work and small coding tasks. Among all the 15 models tested so far it sits at the top of the usable tier, roughly on par with Apple’s on-device models for responsiveness, except this one is open weights. The only practical friction is the custom architecture — stock mlx-serve still rejects it, so I used KRAFTON’s llama.cpp fork. There is an open PR on the mlx-serve repo, once that is merged, it should just work.

Like I hinted in my previous test, I think I will follow up now with a bigger recap, in which I will try to classify all the videos and come up with a workable list. Local AI is becoming something real: you can now do the task of a medium engineer on a local computer, without paying for API tokens and with more than decent results.