{
  "date": "2026-09-23",
  "model": "MiniMax-H3 FL2VA + FastH3 vsa-datafree student",
  "task": "t2va",
  "seed": 1101,
  "steps": 4,
  "geometry": {
    "width": 1344,
    "height": 768,
    "frames": 124,
    "fps": 24,
    "seconds": 5.166666666666667,
    "text_tokens": 457,
    "audio_tokens": 414,
    "video_shape": [
      37,
      24,
      42
    ],
    "prefix": [
      457,
      414
    ],
    "prefix_tiles": 15,
    "video_tiles": 660,
    "topk": 66,
    "compact_rows": 38167,
    "aligned_rows": 38208,
    "local_rows": 4776,
    "padded_rows": 43200,
    "chunk_rows": 1194,
    "kmax": 132,
    "owner_tile_counts": [
      81,
      121,
      132,
      132,
      132,
      132,
      121,
      132
    ],
    "prompt_sha256": "5abb96b31333eea455c241aa17e61d1eaf75d7e78625c80f17302b34ce11fad3"
  },
  "timing_boundary": "HTTP request submission to complete MP4 response bytes; excludes loading, warmup and post-response file validation",
  "sample_policy": "Two warmups, five sequential formal requests per arm; fixed seed and prompt. Published video is formal repeat 1.",
  "arms": [
    {
      "key": "fastest",
      "label": "All-step MXFP8 + Sage",
      "arm": "all_mxfp8",
      "request_seconds": [
        4.668449042132124,
        4.667934790952131,
        4.6809148490428925,
        4.68076366209425,
        4.69352346402593
      ],
      "median_seconds": 4.68076366209425,
      "min_seconds": 4.667934790952131,
      "max_seconds": 4.69352346402593,
      "median_rtf": 0.9059542571795323,
      "under_5": 5,
      "under_playback": 5,
      "warmup_seconds": [
        12.900286988122389,
        4.69785589305684
      ],
      "load_seconds": 142.10446464712732,
      "published_repeat": 1,
      "media": {
        "width": 1344,
        "height": 768,
        "frames": 124,
        "fps": 24,
        "video_seconds": 5.166667,
        "audio_seconds": 5.175,
        "audio_rate": 32000,
        "audio_channels": 2,
        "bytes": 10206291,
        "sha256": "d4901bef7a356fdfc4cc1505629988e4d993db6744f61bf7aeb876ad126f35bd",
        "decode": "PASS"
      },
      "asset_prefix": "benchmarks/t2va-5s-768p"
    },
    {
      "key": "first-step-bf16",
      "label": "First-step BF16 rollback",
      "arm": "first_step_bf16",
      "request_seconds": [
        5.205809125909582,
        5.209279254078865,
        5.220295866951346,
        5.25953430403024,
        5.2396815901156515
      ],
      "median_seconds": 5.220295866951346,
      "min_seconds": 5.205809125909582,
      "max_seconds": 5.25953430403024,
      "median_rtf": 1.0103798452163895,
      "under_5": 0,
      "under_playback": 0,
      "warmup_seconds": [
        15.69795120600611,
        5.226529722800478
      ],
      "load_seconds": 166.11001785215922,
      "published_repeat": 1,
      "media": {
        "width": 1344,
        "height": 768,
        "frames": 124,
        "fps": 24,
        "video_seconds": 5.166667,
        "audio_seconds": 5.175,
        "audio_rate": 32000,
        "audio_channels": 2,
        "bytes": 10141455,
        "sha256": "5dd445775169e2722972e8a6982ce5e61ae7fe233ba46c71e1d3992e2c6f7c15",
        "decode": "PASS"
      },
      "asset_prefix": "benchmarks/t2va-5s-768p"
    }
  ],
  "common": {
    "video_vae": "NVFP4",
    "audio_vae": "native, unchanged",
    "transport": "BF16 RDMA, TP1/SP8",
    "vsa_topk": 66,
    "vsa_video_tiles": 660,
    "flow_shift": 12,
    "audio_flow_shift": 3
  },
  "first_step_bf16": {
    "projections": "original student-fused BF16 tensors retained before quantization",
    "attention": "FlashInfer BF16 VSA, same sparse mask; distinct from the older gallery Triton provider",
    "weight_residency": "first 24 blocks on GPU, last 26 blocks pinned CPU with double-buffered prefetch; transfer included in request timing"
  },
  "caveats": [
    "One prompt and seed; five requests per arm are not a production latency percentile study.",
    "A 5-second request maps to 124 frames, 5.166667 seconds.",
    "The original prompt is unchanged and includes a shot at 6.6 seconds that is beyond this output window.",
    "Both arms use lossy NVFP4 video VAE; first-step rollback does not restore an all-BF16 pipeline.",
    "Quality is presented through original outputs; no subjective quality ranking or all-BF16-reference PSNR is claimed."
  ]
}