Models
18 open-weight models, one page each
Every model on the Green Tokens index has its own page: the index row per GPU, every measured configuration from a running batch of eight upward, H100 against B200 where both were measured, carbon on a spread of grids, and the questions people ask about it. Every figure comes from the same dated snapshot as the index.
Unit watt-hours per million output tokens (Wh/M). GPU only, busy server. Every row is a floor. Updated 27 September 2026.
Lowest index row first
Pick a model
Qwen 3 30B A3B Instruct
15.7Wh/M, 1 × B20030 B parameters, 3 B active; 31 measured configurations.
NVIDIA Nemotron Nano 12B V2
39.4Wh/M, 1 × B20012 B parameters; 14 measured configurations.
NVIDIA Nemotron Nano 9B V2
40.9Wh/M, 1 × B2009 B parameters; 14 measured configurations.
Qwen 3 235B A22B Instruct
66.7Wh/M, 4 × B200235 B parameters, 22 B active; 31 measured configurations.
Llama 4 Scout 17B 16E Instruct
93.4Wh/M, 4 × H100109 B parameters, 17 B active; 20 measured configurations.
Llama 4 Maverick 17B 128E Instruct FP8
134.5Wh/M, 8 × H100400 B parameters, 17 B active; 9 measured configurations.
Qwen 3 235B A22B Instruct FP8
139.1Wh/M, 2 × B200235 B parameters, 22 B active; 23 measured configurations.
Llama 3.1 405B Instruct FP8
318.5Wh/M, 4 × B200405 B parameters; 15 measured configurations.
The figure on each card is the model's lowest index row across the GPUs it was measured on. Closed-weight models served over an API cannot be measured from outside and are not here. Source: ML.ENERGY Leaderboard v3.0 (data compiled 16 February 2026), Text Conversation, 328 stable configurations, retrieved 26 September 2026.