the energy · 16 rows
What a token burns.
Independently measured watt-hours per million output tokens, for the seats research has actually weighed · no extrapolation, no fleet average. Every row prints its measurement source verbatim. Absence is honest: a model without a row has not been measured, which is a fact, not a zero.
Vendored from the released binary · supernovae-st/nika@5d61b7e13f1f (v0.107.0) · digest-verified, re-derived at build, gated in CI
The measured rows
- Qwen/Qwen3-235B-A22B-Thinking-2507176 Wh/Mtok · huggingface · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.6336 J/token (arXiv 2505.06371)
- Qwen/Qwen3-32B86.6 Wh/Mtok · huggingface · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.3116 J/token (arXiv 2505.06371)
- deepseek-ai/DeepSeek-R1660.1 Wh/Mtok · huggingface · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 2.3762 J/token (arXiv 2505.06371)
- deepseek/deepseek-r1660.1 Wh/Mtok · openrouter · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 2.3762 J/token (arXiv 2505.06371)
- nvidia/nvidia-nemotron-nano-9b-v230.3 Wh/Mtok · nvidia · independent-measured
ml.energy leaderboard Dec 2025 · H100 · vLLM · problem-solving · 0.1089 J/token (arXiv 2505.06371)
- openai/gpt-oss-120b11 Wh/Mtok · groq · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.0397 J/token (arXiv 2505.06371)
- openai/gpt-oss-120b11 Wh/Mtok · huggingface · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.0397 J/token (arXiv 2505.06371)
- openai/gpt-oss-120b11 Wh/Mtok · nvidia · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.0397 J/token (arXiv 2505.06371)
- openai/gpt-oss-20b7.9 Wh/Mtok · groq · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.0284 J/token (arXiv 2505.06371)
- openai/gpt-oss-20b7.9 Wh/Mtok · huggingface · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.0284 J/token (arXiv 2505.06371)
- openai/gpt-oss-20b7.9 Wh/Mtok · nvidia · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.0284 J/token (arXiv 2505.06371)
- qwen/qwen3-14b50.7 Wh/Mtok · openrouter · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.1826 J/token (arXiv 2505.06371)
- qwen/qwen3-235b-a22b-thinking-2507176 Wh/Mtok · openrouter · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.6336 J/token (arXiv 2505.06371)
- qwen/qwen3-30b-a3b-thinking-250724.3 Wh/Mtok · openrouter · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.0876 J/token (arXiv 2505.06371)
- qwen/qwen3-32b86.6 Wh/Mtok · groq · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.3116 J/token (arXiv 2505.06371)
- qwen/qwen3-8b48.1 Wh/Mtok · openrouter · independent-measured
ml.energy leaderboard Dec 2025 · B200 · vLLM · problem-solving · 0.1732 J/token (arXiv 2505.06371)