{ "version": "1", "updated": "2026-10-19", "data_version": "2026-10-19", "unit": "Wh per 1,000 queries", "round": "December 2025", "setup": "text generation, 1,000 queries one at a time (batch size 1) on NVIDIA H100 80GB GPUs: one card for class A and B models, several cards for class C; output length fixed at 10 tokens for non-reasoning runs; the score averages ten passes; GPU energy only", "classes": { "A": "fits one consumer GPU", "B": "fits one H100 (single cloud GPU)", "C": "needs more than one H100 (multiple cloud GPUs)" }, "score": "1 to 5 stars within a class for this round: the class's energy range is cut into five equal bands and the lowest band is five stars; stars do not compare across classes", "note": "Only the December 2025 round is included: the benchmark harness changed between rounds (Optimum Benchmark in February 2025, Neuralwatt's AI Energy Benchmarks in December 2025); the project treats watt-hour values as stable across rounds and recalibrates only the stars. A query is a controlled probe with a fixed ten-token response, not a serving workload, so these values are not converted to Wh per million tokens. source_label keeps the model string exactly as the source CSV gives it; model carries the clean Hugging Face id.", "source": { "name": "AI Energy Score leaderboard", "publisher": "AI Energy Score (Hugging Face and Salesforce, with Neuralwatt, Cohere, Meta and Carnegie Mellon University)", "url": "https://huggingface.co/spaces/AIEnergyScore/Leaderboard", "data_url": "https://huggingface.co/spaces/AIEnergyScore/Leaderboard/raw/main/data/energy/text_generation.csv", "documentation": "https://huggingface.github.io/AIEnergyScore/", "version": "December 2025 round", "harness": "Neuralwatt's AI Energy Benchmarks for the December 2025 round; the February 2025 round ran on Optimum Benchmark", "retrieved": "2026-09-26", "licence": "Apache-2.0 as tagged on the Space; the project states no separate licence for the results file, which is reproduced as facts with attribution", "used_for": "GPU energy per 1,000 text-generation queries at batch size 1 on NVIDIA H100 80GB GPUs (one card for class A and B models, several for class C)" }, "rows": [ { "model": "tencent/Hunyuan-1.8B-Instruct", "class": "A", "wh_per_1000_queries": 4.2, "score": 5, "source_label": "tencent/Hunyuan-1.8B-Instruct" }, { "model": "google/gemma-3-270m", "class": "A", "wh_per_1000_queries": 9.3, "score": 5, "source_label": "google/gemma-3-270m" }, { "model": "Qwen/Qwen3-0.6B", "class": "A", "wh_per_1000_queries": 11.8, "score": 5, "source_label": "Qwen/Qwen3-0.6B" }, { "model": "google/gemma-3-1b-pt", "class": "A", "wh_per_1000_queries": 12.3, "score": 5, "source_label": "google/gemma-3-1b-pt" }, { "model": "ibm-granite/granite-4.0-micro", "class": "A", "wh_per_1000_queries": 16.6, "score": 5, "source_label": "ibm-granite/granite-4.0-micro" }, { "model": "microsoft/Phi-4-reasoning-plus", "class": "A", "wh_per_1000_queries": 18.4, "score": 5, "source_label": "microsoft/Phi-4-reasoning-plus" }, { "model": "HuggingFaceTB/SmolLM3-3B", "class": "A", "wh_per_1000_queries": 18.4, "score": 5, "source_label": "HuggingFaceTB/SmolLM3-3B" }, { "model": "swiss-ai/Apertus-8B-Instruct-2509", "class": "A", "wh_per_1000_queries": 21.7, "score": 4, "source_label": "swiss-ai/Apertus-8B-Instruct-2509" }, { "model": "mistralai/Mistral-Nemo-Instruct-2407", "class": "A", "wh_per_1000_queries": 23.2, "score": 4, "source_label": "mistralai/Mistral-Nemo-Instruct-2407" }, { "model": "Qwen/Qwen2.5-Coder-14B", "class": "A", "wh_per_1000_queries": 27.1, "score": 4, "source_label": "Qwen/Qwen2.5-Coder-14B" }, { "model": "google/gemma-3-4b-pt", "class": "A", "wh_per_1000_queries": 34.4, "score": 4, "source_label": "google/gemma-3-4b-pt" }, { "model": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "class": "C", "wh_per_1000_queries": 36.6, "score": 5, "source_label": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B" }, { "model": "Qwen/Qwen3-30B-A3B", "class": "B", "wh_per_1000_queries": 37.4, "score": 5, "source_label": "Qwen/Qwen3-30B-A3B" }, { "model": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "class": "C", "wh_per_1000_queries": 49.5, "score": 5, "source_label": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B" }, { "model": "google/gemma-3-27b-pt", "class": "B", "wh_per_1000_queries": 50.4, "score": 2, "source_label": "google/gemma-3-27b-pt" }, { "model": "LGAI-EXAONE/EXAONE-4.0-32B", "class": "B", "wh_per_1000_queries": 54.5, "score": 1, "source_label": "LGAI-EXAONE/EXAONE-4.0-32B" }, { "model": "ibm-granite/granite-4.0-h-tiny", "class": "A", "wh_per_1000_queries": 81.8, "score": 1, "source_label": "ibm-granite/*granite-4.0-h-tiny", "note": "The source marks this row with an asterisk; model holds the clean Hugging Face id." }, { "model": "zai-org/GLM-4.5-Air", "class": "C", "wh_per_1000_queries": 220.5, "score": 5, "source_label": "zai-org/GLM-4.5-Air" }, { "model": "Qwen/Qwen3-235B-A22B", "class": "C", "wh_per_1000_queries": 305.8, "score": 4, "source_label": "Qwen/Qwen3-235B-A22B" }, { "model": "swiss-ai/Apertus-70B-Instruct-2509", "class": "C", "wh_per_1000_queries": 697.5, "score": 3, "source_label": "swiss-ai/Apertus-70B-Instruct-2509" }, { "model": "zai-org/GLM-4.5", "class": "C", "wh_per_1000_queries": 1225.4, "score": 1, "source_label": "zai-org/GLM-4.5" }, { "model": "zai-org/GLM-4.6", "class": "C", "wh_per_1000_queries": 1273.5, "score": 1, "source_label": "zai-org/GLM-4.6" } ] }