What is an open-weights model?

Why it matters

The Open Source Initiative defines open weights as the final weights and biases of a trained neural network. Publishing them lets others run, study or fine-tune the model on their own hardware. Open weights are not the same as open source AI, which also needs the training code and the training data, or full details of the data.

Because the model files are public, the parameter count, active parameters, quantisation level and memory needs can be read from documentation. Proprietary models do not disclose these, which makes their energy harder to estimate.

An example

An open-weights model can be downloaded as a quantised checkpoint and run on a GPU while its power draw is recorded. A proprietary model can only be called through its provider's API, so the hardware and energy behind each answer stay hidden. Measured open-weights models therefore act as the reference points for estimating the rest.

How Greenpixie uses it

Greenpixie benchmarked 32 open-weights LLMs, from 1.7 billion to 1 trillion total parameters. The set was chosen for:

• quantised checkpoints available on public model hosting platforms
• even coverage across parameter sizes, so the predictions have few gaps
• a mix of providers and specialisations, including coding and reasoning models
• both dense and mixture-of-experts designs

Greenpixie prioritises the largest available open-weights models, as these give the best basis for estimating large proprietary models. Regression models then extend the results to models that cannot be measured directly.