Model weights
Model weights are the numeric parameters a training run produces — the entire learned content of a model — and a published checkpoint of them, together with the code and configuration needed to run it, is enough to reproduce that model's behaviour anywhere the hardware exists.
A checkpoint is less than people expect: tensors in a file, usually safetensors, plus a configuration and a tokeniser. No training data, no training code, no optimiser state. That is why open-weight is the accurate term and open-source generally is not — you can run and serve the model, and you cannot rebuild it.
Their size is arithmetic. Parameter count multiplied by bytes per parameter gives the memory the model occupies before a single request arrives, so a 200-billion-parameter model at 16-bit precision is roughly 400GB resident. That fixed footprint is why serving is a scale business with a threshold under which it does not work at all.
Weights alone do not determine behaviour. Serving precision, the inference stack and the sampling parameters all sit between a checkpoint and an answer, which is how two endpoints running the identical file produce measurably different output while reporting the same model name.
Publication is irreversible, and that is the durability argument for building on open weights. A released checkpoint cannot be recalled, deprecated out from under you or repriced by its author — the copy that exists keeps existing, whatever the publisher does next.
What it costs you
Weights are a fixed cost and tokens are a variable one, and the price of an inference API is essentially the first divided by the second. Memory has to be filled before anything can be served, so cost per token falls as utilisation rises and a model nobody calls is the most expensive thing on a fleet. It is also the mechanism behind open weights being cheaper: the same fixed cost is competed over by several sellers rather than recovered by one, and the research that produced the checkpoint has already been paid for by somebody else.
Related terms
- Open-weight modelAn open-weight model is one whose trained parameters are published under a licence permitting others to download, run and serve them, which is what makes it possible for anyone other than the original lab to offer that model as an API.
- QuantizationQuantization is the compression of a model's weights from higher-precision numbers to lower-precision ones — typically 16-bit down to 8-bit or 4-bit — which reduces the memory and compute needed to serve it, at some cost to output quality.
- Mixture of experts (MoE)A mixture-of-experts model divides its parameters into many specialised sub-networks and routes each token through only a few of them, so the parameters actually used per token — the active parameters — are a small fraction of the model's total size.
Go deeper
Every rate quoted above is published in full on the pricing page, alongside the model publisher’s own official rate.