What's great
I switched to Ollama from clunkier solutions and I have no regrets. Multimodal model support is well-implemented - feeding images through the API is just as easy as text. It’s great that the project is evolving so fast; support for new models usually pops up almost the day after they drop on Hugging Face. This is the simplest way to "play around" with modern AI locally without the headache of setting up the environment.
What needs improvement
I’d love to see a simpler way to import my own .gguf files downloaded outside the official Ollama library, without having to manually define all the parameters in a Modelfile.
How much disk space do common models require locally?
Typically, popular medium-sized models (7B-8B parameters) like Llama 3 or Mistral take up about 4.5-5 GB in standard 4-bit quantization
How simple is creating and customizing your own models?
It's simple using the Modelfile system, the process is very similar to writing a Dockerfile