{
  "date": "2026-05-23",
  "model": "Lasimeri/MiniMax-M2.7-int4-AutoRound",
  "hardware": "4x Intel Arc Pro B70 32GB",
  "os": "Ubuntu 24.04.4 LTS",
  "quality_status": "quality_passed",
  "context_tokens": 2048,
  "benchmark_shape": {
    "input_tokens": 512,
    "output_tokens": 1536,
    "tensor_parallel": 4,
    "max_num_seqs": 1,
    "max_num_batched_tokens": 512
  },
  "mean_total_tokens_per_second": 110.8962457692952,
  "mean_output_tokens_per_second": 83.1721843269714,
  "total_tokens_per_second": [
    109.82634516832029,
    111.29770064350417,
    111.0660964054885,
    111.39484085986787
  ],
  "output_tokens_per_second": [
    82.36975887624021,
    83.47327548262813,
    83.29957230411637,
    83.5461306449009
  ],
  "summary_artifact_on_originating_machine": "/mnt/fast-ai/bench-results/minimax-m27-b70-89tps/strict/minimax-repro-minimax-moe-full-forward-customop-plus-output-ar-strict-tp4-ctx2048-mbt512-bs256-20260523T145201Z-summary.json",
  "notes": [
    "Effective total throughput exceeded the user's >90 tok/s target.",
    "Output-only throughput was below the earlier 89-class and 94-class lanes.",
    "The setup passed strict quality gates before serving was started."
  ]
}