AI Models

Most people meet AI through one chat window run by one company. Open models are now good enough to run on a desk, and they can keep data in-house, cost little to run on hardware you already own and be inspected in detail. However, few teams have tested them on their own work, so the choice is often made on vendor claims.

The difficulty is that each model is good at something different, and a benchmark table rarely says which. A model that writes clean code can be slow, while a fast one can invent an answer it should refuse. A model also changes with its quantisation (the compressed file it runs from), so the same name can give two different results.

I therefore run open models on my own machines, served as GGUF files by llama.cpp, with Claude and Codex beside them in the cloud. Most have a job they earned in testing, such as coding, extraction or reasoning, while two vision models are still on trial. Each card below says what I would use the model for and what I would still test. Models that were downloaded but never run are left out.

I also fine-tune. My Kev checkpoints follow Jared Palmer's open-source recipe on an open base model; the recipe is Palmer's, and the checkpoints are my own. Whatever model does the work, a plain test, not a model, judges the result.

For anyone who has to choose a model, the lesson is to test it on your own work first, and keep it only while it earns its place.