Start with a known family
Well-known families
Start with a model family that many people use, from a recognised publisher: Google's Gemma, Alibaba's Qwen, Mistral, Meta's Llama and Muse models, DeepSeek and others. The apps' built-in model lists and Ollama's library highlight popular ones. Popular models are better tested, better supported and easier to get help with.
Chat or instruct versions
Many models come as a 'base' version and an 'instruct' or 'chat' version. For an assistant, you want the instruct or chat one. Some are also labelled for vision (they can read images), coding, or reasoning (they think step by step before answering, which is slower but better on hard problems).
Licences
Open-weight models come with licences. Some, such as Apache 2.0 or MIT, allow almost any use including commercial. Others add conditions, such as limits for very large companies or acceptable-use rules. For personal use this rarely matters; for business use, read the licence on the model page before you build on it.
Test on your own tasks
Benchmarks tell you little about your work. Run your three real tasks on two candidate models and compare: accuracy, tone, speed, and whether it follows instructions. Keep the winner, delete the other to save disk space, and check again in a few months, because new models arrive constantly.
What three tasks would you most like a private assistant to do well? Write them down; they are your test.