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-source 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.
The benchmark set includes models ranging from 1.7 billion to 1 trillion parameters. Greenpixie prioritises collecting data from the largest available open-source 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. Benchmarking is performed on current-generation GPU architectures to ensure alignment with the hardware used in production cloud AI services.
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 Closed-Source Model Parameter Estimation
Proprietary models do not publish their full architecture specifications. To apply the predictive models described in Section 7, Greenpixie requires estimated values for model characteristics such as total parameter count and active parameter count.
These estimates are derived through a multi-source process that combines third-party research, published performance benchmarks and throughput data, and architectural characteristics of comparable open-source models. For mixture-of-experts architectures, frontier open-source models provide reference points for typical active-to-total parameter ratios. Where a model family has earlier generations with published architecture details, those disclosures serve as anchor points for extrapolation to later versions.
Greenpixie cross-validates parameter estimates against economic plausibility by comparing the predicted kWh-per-token values against the model's commercial token pricing. Where the implied energy cost is inconsistent with the economics of operating the model at published prices, the underlying parameter assumptions are revisited. This process is described further in Section 9.4.
The specific estimation techniques, source prioritisation, and cross-validation thresholds are proprietary. Where the available information is insufficient even for this estimation process, a model may be assigned values based on comparable models with similar pricing and stated capabilities, subject to internal review.
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.