This is our fourth video in the series: Local AI – models that you can actually run on an entry level machine, like a 16GB MacBookPro.

This time we’re testing LFM 2.5 3B, a 3 billion model from Liquid AI, the US neolab that started at MIT.

The testing methodology is the same: for text generation, we’re asking “what is a dolphin?” and for code generation we’re asking for a small utility in Python that can count words in files (and in multiple folders, if asked).

The text generation is very fast and the answer is more than reasonable. No hallucination and good prompt adherence.

Coding is decent. The model spent about 2 minutes on the task and delivered a working Python file. It also asked for tools right after the prompt was processed, and it went on searching the web first. After that, it generated the Python utlity on disk.

All in all, this looks like a solid model. What makes it even more interesting is the fact that it is marketed as an OCR and document recognition model. I didn’t test those features yet, maybe in a future video.

For newcomers: the actual screen recordings are not sped up, or edited in any way, because I want my subscribers to have a real idea about what it is to run those models on their hardware.

If you like this, like, share and subscribe.

If you want other models tested, drop a comment, I read them all.

For more details about how I set up the test environment, as well as the list of the previous tests, check out this playlist: https://www.youtube.com/playlist?list=PLYoyB8nx4G30