Lead Green

Compare a model's energy per million output tokens on H100 and B200. Args: model: model name, slug or Hugging Face id. itl_deadline_ms: optional latency ceiling in milliseconds. If given, each GPU's row is the lowest-energy configuration with median inter-token latency at or under it. If omitted, the index rows are compared. Returns the best row per GPU and the ratio between them. Models measured on only one GPU return that single row.

Energy per million output tokens at every measured running batch for a model on a GPU. Args: model: model name, slug or Hugging Face id. gpu: "H100" or "B200". The running batch (avg_running_batch, vLLM num_requests_running) is how many requests the server decodes at once. Energy per token falls as it grows. Use this to see where a lightly loaded deployment sits relative to the index.

GPU energy per million output tokens for a model, from the index. Args: model: model name, slug or Hugging Face id (for example "Qwen 3 32B", "qwen-3-32b" or "Qwen/Qwen3-32B"). gpu: optional, "H100" or "B200". If omitted, every GPU the model was measured on is returned. itl_deadline_ms: optional latency ceiling in milliseconds. If given, each GPU's row is re-selected from every measured configuration: the lowest energy whose median inter-token latency is at or under it. If omitted, the index rows (the index's own ceiling) are returned. Each row is the most energy-efficient configuration found for that model on that GPU while keeping median inter-token latency under the ceiling, on a text-conversation workload. The response reports the ceiling used.

Estimate GPU energy, and optionally carbon and energy at the meter, for serving a number of output tokens. Args: model: model name, slug or Hugging Face id. output_tokens: how many output tokens will be generated (for example 50_000_000 for fifty million). gpu: "H100" or "B200". If omitted, the most efficient measured GPU is used. country: optional country name or ISO 3166 code (or "World") to convert the energy to kg CO2e using that grid's annual average intensity. running_batch: optional. The average running batch the server will hold: how many requests it decodes at once (vLLM num_requests_running). If given, the estimate uses the measured configuration nearest to it, on the same GPU count as the best row unless num_gpus is set. A small running batch can raise energy per token several-fold. num_gpus: optional. Use configurations on this many GPUs (for example 2 for 2 x H100). itl_deadline_ms: optional latency ceiling in milliseconds. The lowest-energy configuration with median inter-token latency at or under it is used. If omitted, the index's own ceiling applies. pue: optional power usage effectiveness of the facility, at least 1.0. system_power_ratio: optional ratio of whole-server power to GPU power, at least 1.0 (a DGX B200 at its 14.3 kW maximum over 8 x 1.0 kW of GPU power is about 1.79). requests_in_flight: the old name for running_batch, still accepted. Returns energy in Wh and kWh, the configuration used, and, with a country, kg CO2e. With pue or system_power_ratio it also returns energy at the meter, which rests on those assumptions. The measured figures are GPU-only floors; see the caveat.

Manufacturer power ceilings (max TDP), memory, and MIG slice profiles for data-centre GPUs. Args: gpu: optional filter such as "H100", "H200", "A100" or "B200". Omit for all. Whole-system ceilings (for example DGX B200) come with the matching GPU; they help choose a system_power_ratio for estimate.

Annual average grid carbon intensity for a country, in g CO2e per kWh. Args: country: country name (common names such as "UK", "US" or "Turkey" work), ISO 3166 alpha-3 code, alpha-2 code where the data carries it, or "World". display_name is the short name the site shows; name is the data source's label.

List every model on the Green Tokens index with the GPUs it was measured on. Returns names, Hugging Face ids, parameter counts and the GPUs available, so you can pick exact arguments for the other tools.

Every source behind the index, with version, retrieval date and licence, plus the API endpoints.