{
  "version": "1",
  "updated": "2026-09-27",
  "data_version": "2026-09-27",
  "unit": "Wh per million output tokens",
  "itl_deadline_ms": 500,
  "source": {
    "name": "ML.ENERGY Leaderboard",
    "publisher": "The ML.ENERGY Initiative (started at SymbioticLab, University of Michigan)",
    "url": "https://ml.energy/leaderboard/",
    "data_url": "https://raw.githubusercontent.com/ml-energy/leaderboard/master/public/data/tasks/lm-arena-chat.json",
    "version": "3.0 (released 2025-12-01); data compiled 2026-02-16",
    "released": "2025-12-01",
    "compiled": "2026-02-16",
    "compiled_evidence": "the leaderboard's public/data/index.json last_updated, and the last commit to the task files, both 2026-02-16",
    "retrieved": "2026-09-26",
    "licence": "Apache-2.0 per the ml-energy/benchmark-v3 dataset card",
    "licence_evidence": [
      "Hugging Face dataset ml-energy/benchmark-v3: tagged license:apache-2.0, ships a LICENSE file, and its dataset card's License section reads Apache-2.0",
      "The ML.ENERGY Leaderboard web app's footer says Apache 2.0",
      "The leaderboard repository (github.com/ml-energy/leaderboard) carries no licence file; its JSON files are reproduced here as published measurements until the ML.ENERGY team confirms that the dataset licence covers them"
    ],
    "workload": "Text Conversation (lm-arena-chat)",
    "method": "Energy per output token is steady-state GPU energy divided by the output tokens generated in the steady-state window (the benchmark's steady_state_energy_per_token), measured with Zeus under vLLM 0.11.1 on H100 80GB HBM3 and B200 servers. Energy per request is that figure multiplied by the average output length, not a separate measurement. Prompt processing is included.",
    "selection": "Each index row is the lowest-energy configuration for its model and GPU whose median inter-token latency does not exceed 500 ms, the leaderboard's own default ceiling; no measured configuration reaches it.",
    "cite": "Chung et al., The ML.ENERGY Benchmark: Toward Automated Inference Energy Measurement and Optimization, NeurIPS 2025 Datasets and Benchmarks, arXiv 2505.06371",
    "cite_url": "https://arxiv.org/abs/2505.06371",
    "used_for": "index, per-model configurations, concurrency curves, measured GPU power",
    "configurations": "328 stable configurations: the leaderboard publishes the runs that passed its stability filter; every run, including the unstable ones, is in the gated Hugging Face dataset ml-energy/benchmark-v3"
  },
  "caveat": "GPU energy only, measured at steady state on a busy server (the most energy-efficient configuration under a latency limit). Host, network, cooling and idle time are not counted. Every figure is a floor; a lightly used server sits well above it.",
  "boundary": "gpu_only",
  "slug": "qwen-3-14b",
  "name": "Qwen 3 14B",
  "model_id": "Qwen/Qwen3-14B",
  "params_b": 14.0,
  "active_params_b": 14.0,
  "architecture": "Dense Transformer",
  "gpus": [
    "B200",
    "H100"
  ],
  "configurations": [
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 7.9,
      "wh_per_million_output_tokens": 194.25,
      "joules_per_output_token": 0.6993092,
      "median_itl_ms": 7.2,
      "p90_itl_ms": 8.4,
      "p99_itl_ms": 17.9,
      "tokens_per_second": 1086.7,
      "gpu_power_watts_total": 760,
      "avg_output_tokens": 615,
      "under_itl_deadline": true,
      "avg_running_batch": 7.9,
      "p95_itl_ms": 9.7,
      "row_id": "qwen-3-14b/B200/1/8",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 15.9,
      "wh_per_million_output_tokens": 110.6,
      "joules_per_output_token": 0.3981641,
      "median_itl_ms": 7.8,
      "p90_itl_ms": 11.5,
      "p99_itl_ms": 21.0,
      "tokens_per_second": 1912.8,
      "gpu_power_watts_total": 762,
      "avg_output_tokens": 627,
      "under_itl_deadline": true,
      "avg_running_batch": 15.9,
      "p95_itl_ms": 14.7,
      "row_id": "qwen-3-14b/B200/1/16",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 67.81,
      "joules_per_output_token": 0.2441301,
      "median_itl_ms": 9.4,
      "p90_itl_ms": 18.6,
      "p99_itl_ms": 34.4,
      "tokens_per_second": 2866.2,
      "gpu_power_watts_total": 700,
      "avg_output_tokens": 605,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 23.2,
      "row_id": "qwen-3-14b/B200/1/32",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.8,
      "wh_per_million_output_tokens": 46.36,
      "joules_per_output_token": 0.1668923,
      "median_itl_ms": 12.7,
      "p90_itl_ms": 25.5,
      "p99_itl_ms": 45.0,
      "tokens_per_second": 4176.1,
      "gpu_power_watts_total": 697,
      "avg_output_tokens": 609,
      "under_itl_deadline": true,
      "avg_running_batch": 63.8,
      "p95_itl_ms": 30.7,
      "row_id": "qwen-3-14b/B200/1/64",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.7,
      "wh_per_million_output_tokens": 34.02,
      "joules_per_output_token": 0.1224768,
      "median_itl_ms": 19.1,
      "p90_itl_ms": 34.6,
      "p99_itl_ms": 54.3,
      "tokens_per_second": 5786.6,
      "gpu_power_watts_total": 709,
      "avg_output_tokens": 606,
      "under_itl_deadline": true,
      "avg_running_batch": 127.7,
      "p95_itl_ms": 40.3,
      "row_id": "qwen-3-14b/B200/1/128",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 255.6,
      "wh_per_million_output_tokens": 24.77,
      "joules_per_output_token": 0.0891861,
      "median_itl_ms": 21.2,
      "p90_itl_ms": 32.6,
      "p99_itl_ms": 53.2,
      "tokens_per_second": 10708.7,
      "gpu_power_watts_total": 955,
      "avg_output_tokens": 624,
      "under_itl_deadline": true,
      "avg_running_batch": 255.6,
      "p95_itl_ms": 37.4,
      "row_id": "qwen-3-14b/B200/1/256",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 512,
      "avg_requests_in_flight": 511.3,
      "wh_per_million_output_tokens": 21.4,
      "joules_per_output_token": 0.0770223,
      "median_itl_ms": 39.1,
      "p90_itl_ms": 56.1,
      "p99_itl_ms": 71.6,
      "tokens_per_second": 11858.0,
      "gpu_power_watts_total": 913,
      "avg_output_tokens": 621,
      "under_itl_deadline": true,
      "avg_running_batch": 511.3,
      "p95_itl_ms": 60.0,
      "row_id": "qwen-3-14b/B200/1/512",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 1024,
      "avg_requests_in_flight": 1022.0,
      "wh_per_million_output_tokens": 17.21,
      "joules_per_output_token": 0.0619441,
      "median_itl_ms": 61.5,
      "p90_itl_ms": 78.1,
      "p99_itl_ms": 150.3,
      "tokens_per_second": 15050.1,
      "gpu_power_watts_total": 932,
      "avg_output_tokens": 613,
      "under_itl_deadline": true,
      "avg_running_batch": 1022.0,
      "p95_itl_ms": 85.3,
      "row_id": "qwen-3-14b/B200/1/1024",
      "repeat": 1
    },
    {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 2048,
      "avg_requests_in_flight": 2044.6,
      "wh_per_million_output_tokens": 15.21,
      "joules_per_output_token": 0.0547711,
      "median_itl_ms": 112.6,
      "p90_itl_ms": 142.4,
      "p99_itl_ms": 215.1,
      "tokens_per_second": 15875.0,
      "gpu_power_watts_total": 869,
      "avg_output_tokens": 613,
      "under_itl_deadline": true,
      "avg_running_batch": 2044.6,
      "p95_itl_ms": 153.5,
      "row_id": "qwen-3-14b/B200/1/2048",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 8,
      "avg_requests_in_flight": 8.0,
      "wh_per_million_output_tokens": 196.14,
      "joules_per_output_token": 0.7060932,
      "median_itl_ms": 12.5,
      "p90_itl_ms": 13.2,
      "p99_itl_ms": 14.4,
      "tokens_per_second": 627.4,
      "gpu_power_watts_total": 443,
      "avg_output_tokens": 620,
      "under_itl_deadline": true,
      "avg_running_batch": 8.0,
      "p95_itl_ms": 13.4,
      "row_id": "qwen-3-14b/H100/1/8",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 16,
      "avg_requests_in_flight": 16.0,
      "wh_per_million_output_tokens": 110.63,
      "joules_per_output_token": 0.3982738,
      "median_itl_ms": 13.3,
      "p90_itl_ms": 14.1,
      "p99_itl_ms": 17.9,
      "tokens_per_second": 1168.2,
      "gpu_power_watts_total": 465,
      "avg_output_tokens": 622,
      "under_itl_deadline": true,
      "avg_running_batch": 16.0,
      "p95_itl_ms": 14.6,
      "row_id": "qwen-3-14b/H100/1/16",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 32,
      "avg_requests_in_flight": 31.9,
      "wh_per_million_output_tokens": 67.81,
      "joules_per_output_token": 0.2441004,
      "median_itl_ms": 15.0,
      "p90_itl_ms": 17.2,
      "p99_itl_ms": 28.4,
      "tokens_per_second": 2023.2,
      "gpu_power_watts_total": 494,
      "avg_output_tokens": 611,
      "under_itl_deadline": true,
      "avg_running_batch": 31.9,
      "p95_itl_ms": 17.8,
      "row_id": "qwen-3-14b/H100/1/32",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 64,
      "avg_requests_in_flight": 63.8,
      "wh_per_million_output_tokens": 46.44,
      "joules_per_output_token": 0.1671736,
      "median_itl_ms": 18.9,
      "p90_itl_ms": 21.7,
      "p99_itl_ms": 45.0,
      "tokens_per_second": 3140.3,
      "gpu_power_watts_total": 525,
      "avg_output_tokens": 617,
      "under_itl_deadline": true,
      "avg_running_batch": 63.8,
      "p95_itl_ms": 23.1,
      "row_id": "qwen-3-14b/H100/1/64",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 96,
      "avg_requests_in_flight": 95.9,
      "wh_per_million_output_tokens": 39.87,
      "joules_per_output_token": 0.1435253,
      "median_itl_ms": 23.5,
      "p90_itl_ms": 26.2,
      "p99_itl_ms": 60.4,
      "tokens_per_second": 3774.8,
      "gpu_power_watts_total": 542,
      "avg_output_tokens": 623,
      "under_itl_deadline": true,
      "avg_running_batch": 95.9,
      "p95_itl_ms": 29.9,
      "row_id": "qwen-3-14b/H100/1/96",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 128,
      "avg_requests_in_flight": 127.7,
      "wh_per_million_output_tokens": 35.52,
      "joules_per_output_token": 0.1278583,
      "median_itl_ms": 26.2,
      "p90_itl_ms": 30.0,
      "p99_itl_ms": 72.4,
      "tokens_per_second": 4474.4,
      "gpu_power_watts_total": 572,
      "avg_output_tokens": 614,
      "under_itl_deadline": true,
      "avg_running_batch": 127.7,
      "p95_itl_ms": 36.9,
      "row_id": "qwen-3-14b/H100/1/128",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 192,
      "avg_requests_in_flight": 191.7,
      "wh_per_million_output_tokens": 32.06,
      "joules_per_output_token": 0.1154279,
      "median_itl_ms": 33.4,
      "p90_itl_ms": 39.2,
      "p99_itl_ms": 90.9,
      "tokens_per_second": 5309.1,
      "gpu_power_watts_total": 613,
      "avg_output_tokens": 613,
      "under_itl_deadline": true,
      "avg_running_batch": 191.7,
      "p95_itl_ms": 47.3,
      "row_id": "qwen-3-14b/H100/1/192",
      "repeat": 1
    },
    {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 253.4,
      "wh_per_million_output_tokens": 31.09,
      "joules_per_output_token": 0.1119169,
      "median_itl_ms": 38.4,
      "p90_itl_ms": 49.6,
      "p99_itl_ms": 137.3,
      "tokens_per_second": 5760.3,
      "gpu_power_watts_total": 645,
      "avg_output_tokens": 625,
      "under_itl_deadline": true,
      "avg_running_batch": 253.4,
      "p95_itl_ms": 67.3,
      "row_id": "qwen-3-14b/H100/1/256",
      "repeat": 1
    }
  ],
  "best": {
    "B200": {
      "gpu": "B200",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 2048,
      "avg_requests_in_flight": 2044.6,
      "wh_per_million_output_tokens": 15.21,
      "joules_per_output_token": 0.0547711,
      "median_itl_ms": 112.6,
      "p90_itl_ms": 142.4,
      "p99_itl_ms": 215.1,
      "tokens_per_second": 15875.0,
      "gpu_power_watts_total": 869,
      "avg_output_tokens": 613,
      "under_itl_deadline": true,
      "avg_running_batch": 2044.6,
      "p95_itl_ms": 153.5,
      "row_id": "qwen-3-14b/B200/1/2048",
      "repeat": 1
    },
    "H100": {
      "gpu": "H100",
      "num_gpus": 1,
      "tensor_parallel": 1,
      "data_parallel": 1,
      "expert_parallel": 1,
      "precision": "BF16",
      "max_num_seqs": 256,
      "avg_requests_in_flight": 253.4,
      "wh_per_million_output_tokens": 31.09,
      "joules_per_output_token": 0.1119169,
      "median_itl_ms": 38.4,
      "p90_itl_ms": 49.6,
      "p99_itl_ms": 137.3,
      "tokens_per_second": 5760.3,
      "gpu_power_watts_total": 645,
      "avg_output_tokens": 625,
      "under_itl_deadline": true,
      "avg_running_batch": 253.4,
      "p95_itl_ms": 67.3,
      "row_id": "qwen-3-14b/H100/1/256",
      "repeat": 1
    }
  }
}