Open-weight model
An 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.
Open-weight is a narrower claim than open-source. The weights are released; the training data, the training code and the full recipe usually are not. That is enough to run and serve the model, and not enough to reproduce it, which is why the more precise term has largely displaced the looser one.
The licences differ in ways that matter commercially. MIT and Apache 2.0 impose essentially no restriction on serving or embedding the model in a product. Others add acceptable-use terms, attribution requirements, or thresholds above which separate permission is needed. Every model in this catalogue carries its licence on its own page for that reason.
The distinction from a closed-weight model is structural, not ideological. A closed-weight model has exactly one seller, so its price is whatever that seller decides. Open weights can be served by anyone with the hardware, so the price converges on the cost of the compute plus a margin.
That structure is the entire basis of this service, and it is also why we do not resell closed-weight models: no licence to do so exists, and reselling access to one generally breaches the origin provider's terms.
What it costs you
Open weights turn inference into a market with several sellers instead of a single vendor's price list. That is the mechanism behind every rate on this site being below the model publisher's own official rate — not a promotional discount, but a different cost structure, since we are paying for GPU time rather than for the research that produced the model. It also removes the pricing risk that comes with a single supplier, because the weights you build on cannot be withdrawn or repriced out from under you.
Worked from today’s rates
GLM-5.2 is published by Z.ai under the MIT licence. Because those weights are downloadable and servable, the same model is available here at $0.82/M input against Z.ai's official $1.40/M.
Related terms
- Inference providerAn inference provider is a company that runs open-weight model weights on its own or rented GPUs and sells access to them over an API, as distinct from the lab that trained and published the model.
- 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.