From votes to a leaderboard

How is the energy impact of the models calculated?

The arena uses the methodology developed by Ecologits (GenAI Impact) to provide an estimate of the energy footprint associated with inferring conversational generative AI models. This estimate allows users to compare the environmental impact of different AI models for the same query. This transparency is essential to encourage the development and adoption of more eco-friendly AI models.

Ecologits applies the principles of life cycle assessment (LCA) in accordance with ISO 14044, focusing for the moment on the impact of inference (i.e., the use of models to answer queries) and the manufacturing of graphics cards (resource extraction, manufacturing and transport).

The model's power consumption is estimated by taking into account various parameters such as the size and architecture of the AI model used, the location of the servers where the models are deployed, and the number of output tokens. The calculation of the global warming potential indicator, expressed in CO2 equivalent, is derived from the measurement of the model's power consumption.

It is important to note that methodologies for assessing the environmental impact of AI are still under development.

Chart data in table form

Updated on 6/16/2026

Download data

Search for a model Search

From votes to a leaderboard

Model Sort	BT score of satisfaction Tooltip Estimated statistical score based on the Bradley-Terry model, reflecting the probability that one model is preferred over another. This score is calculated from all user votes and reactions. For more information, visit the methodology tab. Sort	Average energy (per 1000 tokens) Tooltip Measured in watt-hours, energy consumption represents the electricity used by the model to process a request and generate the corresponding response. Model energy consumption depends on the model's size and architecture. We have chosen to display proprietary models for which we do not have transparent information about their size and architecture in gray as "unanalyzed" (N/A). Sort	Size (parameters) Tooltip Model size in billions of parameters, categorized into five classes. For proprietary models, this size is not reported. Sort	Architecture Tooltip The architecture of an LLM model refers to the design principles that define how the components of a neural network are arranged and interact to transform input data into predictive outputs, including the activation mode of parameters (dense vs. sparse), the specialization of components, and the information processing mechanisms (transformers, convolutional networks, hybrid architectures). Sort	Organization Sort	Licence Sort
lfm2-8b-a1b	976	81 mWh	XS - 8 billion	MoE	Liquid	Open-weight
lfm2-24b-a2b	1013	82 mWh	S - 24 billion	MoE	Liquid	Open-weight
gpt-oss-20b	967	83 mWh	S - 21 billion	MoE	OpenAI	Open-weight
qwen3-30b-a3b	967	83 mWh	S - 30 billion	MoE	Alibaba	Open-weight
phi-3.5-mini-instruct	800	83 mWh	XS - 3.8 billion	Dense	Microsoft	Open-weight
gemma-4-26b-a4b-it	1114	84 mWh	S - 26 billion	MoE	Google	Open-weight
gemma-3-4b	1030	84 mWh	XS - 4 billion	Dense	Google	Open-weight
gemma-3n-e4b-it	1020	84 mWh	XS - 8 billion	Matformer	Google	Open-weight
qwen2.5-7b-instruct	873	88 mWh	XS - 7 billion	Dense	Alibaba	Open-weight
chocolatine-2-14b-instruct-v2.0.3-q8	717	88 mWh	XS - 14 billion	Dense	jpacifico	Open-weight
qwen2-7b-instruct	673	88 mWh	XS - 7 billion	Dense	Alibaba	Open-weight
granite-4.1-8b	980	89 mWh	XS - 8 billion	Dense	IBM	Open-weight
aya-expanse-8b	929	89 mWh	XS - 8 billion	Dense	Cohere	Open-weight
llama-3.1-8b	872	89 mWh	XS - 8 billion	Dense	Meta	Open-weight
gemma-2-9b-it	905	90 mWh	XS - 9 billion	Dense	Google	Open-weight
Yi-1.5-9B-Chat	709	90 mWh	XS - 9 billion	Dense	01-ai	Open-weight
gemma-3-12b	1063	94 mWh	XS - 12 billion	Dense	Google	Open-weight
mistral-nemo-2407	773	94 mWh	XS - 12 billion	Dense	Mistral AI	Open-weight
phi-4	921	96 mWh	XS - 14 billion	Dense	Microsoft	Open-weight
chocolatine-14b-instruct-dpo-v1.2-q4	695	96 mWh	XS - 14 billion	Dense	jpacifico	Open-weight
EuroLLM-22B-Instruct-2512	994	106 mWh	S - 22 billion	Dense	EuroLLM	Open-weight
mistral-small-2506	1059	109 mWh	S - 24 billion	Dense	Mistral AI	Open-weight
magistral-small-2506	1040	109 mWh	S - 24 billion	Dense	Mistral AI	Open-weight
mistral-small-3.1-24b	983	109 mWh	S - 24 billion	Dense	Mistral AI	Open-weight
mistral-small-24b-instruct-2501	955	109 mWh	S - 24 billion	Dense	Mistral AI	Open-weight
gemma-3-27b	1091	112 mWh	S - 27 billion	Dense	Google	Open-weight
gemma-2-27b-it-q8	957	112 mWh	S - 27 billion	Dense	Google	Open-weight
gemma-4-31b-it	1107	117 mWh	S - 31 billion	Dense	Google	Open-weight
qwen3-32b	1030	118 mWh	S - 32 billion	Dense	Alibaba	Open-weight
aya-expanse-32b	960	118 mWh	S - 32 billion	Dense	Cohere	Open-weight
qwq-32b	905	118 mWh	S - 32 billion	Dense	Alibaba	Open-weight
qwen2.5-32b-instruct	904	118 mWh	S - 32 billion	Dense	Alibaba	Open-weight
olmo-3-32b-think	862	118 mWh	S - 32 billion	Dense	Ai2	Open source
qwen3.5-35b-a3b	1002	166 mWh	S - 35 billion	MoE	Alibaba	Open-weight
mixtral-8x7b-instruct-v0.1	824	193 mWh	S - 56 billion	MoE	Mistral AI	Open-weight
qwen3-coder-next	1106	332 mWh	M - 80 billion	MoE	Alibaba	Open-weight
gpt-oss-120b	1019	342 mWh	L - 117 billion	MoE	OpenAI	Open-weight
mistral-small-2603	1104	347 mWh	L - 119 billion	MoE	Mistral AI	Open-weight
nemotron-3-super-120b-a12b	1004	376 mWh	L - 120 billion	MoE	Nvidia	Open-weight
llama-4-scout	992	400 mWh	L - 109 billion	MoE	Meta	Open-weight
llama-3.1-nemotron-70b-instruct	1036	658 mWh	M - 70 billion	Dense	Nvidia	Open-weight
Apertus-70B-Instruct-2509	962	658 mWh	M - 70 billion	Dense	Swiss AI	Open source
llama-3.3-70b	959	658 mWh	M - 70 billion	Dense	Meta	Open-weight
llama-3.1-70b	929	658 mWh	M - 70 billion	Dense	Meta	Open-weight
hermes-4-70b	899	658 mWh	M - 70 billion	Dense	Nous	Open-weight
deepseek-r1-distill-llama-70b	898	658 mWh	M - 70 billion	Dense	DeepSeek	Open-weight
minimax-m2	1016	733 mWh	L - 230 billion	MoE	MiniMax	Open-weight
minimax-m2.5	1013	733 mWh	L - 229 billion	MoE	MiniMax	Open-weight
minimax-m2.7	1001	733 mWh	L - 230 billion	MoE	MiniMax	Open-weight
command-a	1053	857 mWh	L - 111 billion	Dense	Cohere	Open-weight
mixtral-8x22b-instruct-v0.1	758	1063 mWh	L - 176 billion	MoE	Mistral AI	Open-weight
trinity-large-thinking	1109	1524 mWh	L - 398 billion	MoE	Arcee	Open-weight
deepseek-v4-flash	1058	1524 mWh	L - 284 billion	MoE	DeepSeek	Open-weight
trinity-large-preview	970	1524 mWh	L - 398 billion	MoE	Arcee	Open-weight
qwen3.5-397b-a17b	1035	1601 mWh	L - 397 billion	MoE	Alibaba	Open-weight
llama-maverick	1008	1601 mWh	XL - 400 billion	MoE	Meta	Open-weight
glm-4.5	1119	1892 mWh	L - 355 billion	MoE	Zhipu	Open-weight
glm-4.6	1090	1892 mWh	L - 357 billion	MoE	Zhipu	Open-weight
glm-4.7	1005	1892 mWh	L - 357 billion	MoE	Zhipu	Open-weight
Qwen3-Coder-480B-A35B-Instruct	1055	1951 mWh	XL - 480 billion	MoE	Alibaba	Open-weight
kimi-k2.6	1078	3785 mWh	XL - 1000 billion	MoE	Moonshot AI	Open-weight
kimi-k2.5	1065	3785 mWh	XL - 1000 billion	MoE	Moonshot AI	Open-weight
kimi-k2-thinking	1059	3785 mWh	XL - 1000 billion	MoE	Moonshot AI	Open-weight
kimi-k2	1052	3785 mWh	XL - 1000 billion	MoE	Moonshot AI	Open-weight
deepseek-v3-0324	1105	3979 mWh	XL - 685 billion	MoE	DeepSeek	Open-weight
deepseek-v3-chat	1094	3979 mWh	XL - 671 billion	MoE	DeepSeek	Open-weight
deepseek-chat-v3.1	1093	3979 mWh	XL - 685 billion	MoE	DeepSeek	Open-weight
deepseek-r1-0528	1068	3979 mWh	XL - 685 billion	MoE	DeepSeek	Open-weight
DeepSeek-V3.2	1058	3979 mWh	XL - 685 billion	MoE	DeepSeek	Open-weight
deepseek-r1	1023	3979 mWh	XL - 671 billion	MoE	DeepSeek	Open-weight
glm-5.1	1022	4095 mWh	XL - 744 billion	MoE	Zhipu	Open-weight
glm-5	1016	4095 mWh	XL - 744 billion	MoE	Zhipu	Open-weight
mistral-large-2512	1137	4134 mWh	XL - 675 billion	MoE	Mistral AI	Open-weight
deepseek-v4-pro	1058	8890 mWh	XL - 1600 billion	MoE	DeepSeek	Open-weight
llama-3.1-405b	915	9134 mWh	XL - 405 billion	Dense	Meta	Open-weight
hermes-3-llama-3.1-405b	891	9134 mWh	XL - 405 billion	Dense	Nous	Open-weight