What are Open Weights?
Open-weight models, told as a story about a parrot whose every number is published, so anyone can bring it home to their own perch.
There are two ways to use a parrot: ask through their counter, or bring it home to our perch.Ask at the counter and the parrot stays at their house; you pay beans every time. Bring it home and the parrot lives with you.
Either way has worries: borrowing means bean bills and secret beans; bringing home means the perch and the caring.A borrowed parrot costs beans every time, and your secret beans go to their house. Your own parrot needs a pricey perch (GPU) and someone to care for it every day.
Open weights: a parrot whose every number in its head is published, so anyone can put it on their own perch.A parrot has billions of numbers (weights) in its head. Publish them all and anyone can download the bird and put it on their own perch — Llama, Mistral, Qwen are such birds. The very smartest parrots are usually met only at the counter.
Borrowed and own parrots differ in cost, freshness, where data goes, tweaking, and upkeep.Your own parrot can be drilled or made to count beans coarsely as you like. In return, its cost and speed depend on your perch.
Open is neither free nor open source. The numbers are public, the book list isn't, and business use has conditions.The perch and the caring cost money — compare it with the bean bill. If you brought it home to keep secret beans safe, the daily caring is yours too.
Open weights = a parrot whose numbers (weights) are all published, so anyone can put it on their own perch. Secret beans stay home and tweaking is free, but the perch and the caring are yours. Open is neither free nor open source — read the scroll.
A model whose trained parameters (weights) are published so anyone can download it and run inference, fine-tune, or quantize it on their own servers (Llama, Mistral, Qwen, etc.). API models run on the provider's servers, bill per token, and see your data. Unlike true open source, training data and code are usually withheld, and licenses vary on commercial use, user thresholds and attribution. Self-hosting brings GPU cost and operations (LLMOps) with it.
When grown-ups say it
- Open weights
- The parrot with its numbers published. Anyone can download it and put it on their own perch. Llama, Mistral, Qwen.
- Weights
- The numbers in the parrot's head. Billions of numbers, set while reading the books. They are the parrot.
- API model
- The borrowed parrot at the counter. Lives at their house; you pay beans per question. → the bean bill
- Self-hosting
- Putting the parrot on our own perch. Buy the perch (GPU) and care for it daily. → caring for the parrot
- License
- The conditions on the scroll. Whether business use is allowed, whether big firms need permission, whether to credit the name — differs per bird.
- Open weights vs open source
- Numbers open vs books and recipe open too. Open weights publish only the numbers. Which books it read and how it was taught usually stay secret.
- Inference server
- What makes it answer on our perch. We set its cost and speed. → what answering costs and takes
- Quantization
- Counting beans coarsely. Possible because it is your bird — it fits a smaller perch. → coarse counting