6. Eligibility, Selection, and Inclusion Logic
What is included and how it is selected and mapped.
6.1 Model Eligibility for Benchmarking
Models selected for physical benchmarking are open-weights LLMs with quantised checkpoints available on public model hosting platforms. The selection is optimised for two objectives. The first is data density across the parameter scale, ensuring minimal gaps between parameter counts so that the regression models have continuous coverage. The second is provider and task diversity, covering models from multiple providers and spanning different specialisations including general-purpose, coding, and reasoning models, and both dense models and MoE models.
The benchmark set comprises 32 models ranging from 1.7 billion to 1 trillion total parameters. Greenpixie prioritises collecting data from the largest available open-weights models, as these provide the strongest basis for extrapolation to large proprietary models where direct measurement is not possible.
6.2 Quantisation and Hardware Configurations
Benchmarked models are quantised at precision levels that reflect real-world, energy-efficient hyperscaler deployments, covering FP8, INT8, and NVFP4. Benchmarking is performed on current-generation GPU architectures spanning two hardware generations, to ensure alignment with the hardware used in production cloud AI services and to allow the effect of hardware generation on energy per token to be measured directly.
6.3 Inclusion in the Prediction Database
Models are included in Greenpixie's prediction database where sufficient publicly available information exists to determine the model characteristics required by the regression models. Models are added on an ongoing basis as cloud providers expand their catalogues.
6.4 Proprietary Model Estimation
Proprietary models do not publish their architecture specifications, so the model characteristics required by the regression models described in Section 7 cannot be taken from model documentation. In place of a single estimated specification, the methodology establishes the range of specifications a model could plausibly have and reports the resulting energy as a range.
Each proprietary model is assigned to a parameter band, being a bounded range of total parameter counts, on the basis of the tier label within its identifier. Every major provider applies a broadly consistent ordering within its model family, and that ordering is a reliable indication of relative model size. Band boundaries are calibrated against the published distribution of frontier open-weights releases and against independent datasets of notable model parameter counts.
Within a band, the characteristics that cannot be observed are not fixed to a single value. Total parameter count, the proportion of parameters active per request, the deployed quantisation level, the hardware in use, the level of request parallelism, and the training energy uplift are each treated as a range. The prediction is produced by sampling across those ranges, yielding a median and an interquartile range. Section 7.6 describes this process.
Band assignment is the largest single influence on the predicted energy of an individual proprietary model, as the bands span approximately an order of magnitude. Ranges are therefore reported in place of point values, and band assignment is subject to internal review where a model does not follow the usual naming convention of its provider.
The band boundaries, the naming conventions used to assign them, the specifications of the sampled ranges, and the review criteria are proprietary. Band ranges are recalibrated periodically, as the parameter count implied by a given tier label moves as the frontier advances. Compute-optimal scaling research by Hoffmann et al. indicates diminishing returns from raw parameter count, so recalibration reviews both upward movement in the bands and the possibility that the upper bands stabilise.
6.5 Embedding Model Treatment
Embedding models are not directly benchmarked. Their energy values are estimated by matching each embedding model to an LLM with a comparable price point, on the basis that similarly priced models tend to have similar inference resource requirements. Embedding token energy is treated as comparable to input token energy, since LLM prefill and embedding generation are computationally analogous operations.