Choose a model, not just a size

Match the task and language

A bigger parameter count does not guarantee better results for your task. Compare candidates using the same examples in the languages, formats, and domain you actually need. Check the publisher’s evaluations, then run your own small evaluation.

Text generation, embeddings, image generation, and adapters solve different problems. An embedding model does not replace a chat model, and a LoRA adapter normally needs its compatible base model.

Read the model card before downloading

The model card is where publishers document intended use, training information, limitations, evaluations, and often licensing. Missing information is uncertainty, not permission or proof of quality.

• Verify the original publisher and any declared base model.
• Check the license itself, including commercial-use and redistribution terms; an open download is not necessarily an unrestricted license.
• Confirm your runtime supports both the architecture and the exact quantization format.
• Check whether access requires accepting conditions or authenticating to the repository.

Make a useful shortlist

Compare two or three models at the same context and memory capacity. Keep architecture-dependent caveats visible. A model that fits with headroom and gives reliable answers may be more useful than one that is constantly offloading.

Downloads and likes help discovery, but are not quality measurements. Record failure cases as carefully as successes.