when reading about open-source ai, terms like model, architecture, and weights get thrown around like they mean the same thing. they don’t.
here is the simplest way i think about it - building an ai is like building and furnishing a house.
the house analogy
- architecture - the blueprint. the structural design; how many rooms (layers) and how they connect. it defines the potential but holds no actual content. the transformer is the common one today.
- weights - the furnishings. billions of numbers that fill the structure - the ai’s actual knowledge and learned behavior.
- model - the finished house. architecture + trained weights, working together to take an input and give you an output.
a checkpoint is just a snapshot of the weights saved mid-training (a .safetensors or .pt file) so progress isn’t lost. confusingly, released models are often called checkpoints too - which is exactly why people mix up the terms.
why “open-weights” matters
when a company releases an open-weights model - think gemma, llama, mistral, or qwen - they hand you the massive file of trained numbers.
if they only shared the architecture, you’d have empty code - training those numbers from scratch costs millions in compute. sharing the weights means you can download it, run powerful ai on your own hardware, and fine-tune it for your own tasks.
contrast that with a closed model like gpt-4 or gemini - the weights stay locked away, and you only ever reach them through an api.
one caveat - open-weights isn’t the same as fully open-source. the training data and code usually stay private.
so; architecture is the shape, weights are the intelligence, and the model is the two together. simple as that.