For the 12th episode of my series in which I test AI models that fit locally on a 16GB MacBook Pro, I chose K2 Horizon 3.7B. The model comes from the Institute of Foundation Models at Mohammed bin Zayed University of Artificial Intelligence. The lab is fairly new, launched in 2025, and focuses on open-weight models.

The 4-bit version is small, in the 2–3 GB range and it uses about 4 GB of RAM once loaded, which leaves a lot of room for other apps on a 16 GB machine. It’s a dense model, not a mixture-of-experts, and it has a 512k context window. The license is Apache 2.0, and they release way more than just the final weights — training recipes, the inference engine, and intermediate checkpoints are also available.

On a basic text question (“what is a dolphin”) it was quick and the answer was solid, better than average for a model this size. On a small coding task (write a utility that counts words in files in a folder) it called tools multiple times and ended up with usable output. It doesn’t show internal thinking steps, so you only see the final result and any tool use. Still, the code worked after it fixed an issue on its own.

The 7B Model Was Surprisingly Slow

It’s worth mentioning that initially I want to test the 7B version of the same K2 Horizon family. The output quality was good, way above average, but the coding task took roughly 11 minutes. I have a hard rule that if a generation takes more than 10 minutes, I don’t include the model in the tests.

The 3.7B is a very practical choice: small enough to run comfortably on a 16 GB Mac, fast enough for real tasks, and capable enough for text generation and lighter coding work. If course, do not expect it to be a replacement for bigger models on complex projects, but it’s more than usable for the things that actually fit in this size range.

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