{
  "model": "MiniMax-H3-Image",
  "source": {
    "description": "User-supplied MiniMax-H3-Image blog package, blog.md and chart-sources/*.svg",
    "articleSha256": "b24ac40b85ce1e0e1ec63d40928d6d277e12fc016570ac8cf96b31d9baf809ba",
    "charts": [
      { "file": "source-data/02-memory.svg", "sha256": "d54091112078f1830f6a740645e204d5cd27a76b71b3b4c4c1216ed93ae932bf" },
      { "file": "source-data/03-pipeline.svg", "sha256": "aa5fa8d9706485bcd61c6584f2b8934b093f8012debe06e0cdb5ea34b7982ff6" },
      { "file": "source-data/04-times.svg", "sha256": "2c1726a3609a5c46aa23de9bc497be5f5b13422dfe3beaf2a4c83f0747cb8035" }
    ],
    "scope": "Supplied measurements and rounded chart values, not independently rerun. Labels corrected after checking the upstream model and released runtime: 68.7 GB is the BF16 image checkpoint including VAE; timing covers the native model pipeline through PNG. No additional comparisons.",
    "verifiedAgainst": [
      "https://huggingface.co/MiniMaxAI/MiniMax-H3-Image#quick-start",
      "https://github.com/Eliovp-BV/paiton-vllm-plugin/blob/main/models/MiniMax-H3-Image/paiton_h3i/server.py#L92",
      "https://github.com/Eliovp-BV/paiton-vllm-plugin/blob/main/models/MiniMax-H3-Image/results/container-e2e-gptq-20261111.jsonl",
      "https://huggingface.co/EliovpAI/MiniMax-H3-Image-FP8-Paiton-RDNA4/blob/main/MODIFICATIONS.md"
    ]
  },
  "device": "1 × Radeon AI PRO R9700",
  "memory": {
    "unit": "GB",
    "capacityGB": 32,
    "bf16Weights": { "imageCheckpointGB": 68.7, "textEncoderGB": 52.0, "totalGB": 120.7 },
    "paitonPeak": { "transformerGB": 20.3, "otherGpuGB": 9.3, "totalGB": 29.6, "requestMP": 16 },
    "otherGpuAllocations": ["activations", "VAE", "resident text-encoder layers"],
    "excludedPinnedHostGB": 24.4,
    "scope": "Estimated BF16 working weights and observed Paiton peak on-card allocation are different quantities, not a matched memory benchmark. The BF16 image checkpoint includes the VAE; text-encoder weights cover its first 50 layers. GB denotes decimal gigabytes; 27.6 GiB is approximately 29.6 GB."
  },
  "pipeline": {
    "implementation": "Native HIP and C++; no PyTorch or Triton",
    "maxReferenceImages": 3,
    "textEncoder": { "model": "Qwen3-VL", "layers": 50, "weights": "FP8", "pinnedHostGB": 24.4, "runsPerImage": 1 },
    "transformer": { "steps": 6, "residentParametersB": 19.3, "blocks": 50, "weightsAndActivations": "FP8", "rounding": "GPTQ", "adaLN": "Precomputed tables", "attentionQK": "INT8", "attentionPV": "FP8", "tokensAt16MP": 16400, "hipGraphsPerStep": 1 },
    "vaeDecoder": { "weights": "INT8", "tilePixels": 256, "architecture": "ViT", "layers": 24 },
    "output": "PNG",
    "referenceImagePath": "Reference images also pass through the VAE encoder; this branch is not drawn."
  },
  "timing": {
    "unit": "seconds",
    "scope": "Warm native model pipeline through PNG, including output resize; preprocessing, queueing and response transfer excluded",
    "powerLimitWatts": 300,
    "steps": 6,
    "rows": [
      { "mode": "generation", "width": 2048, "height": 2048, "references": 0, "seconds": 6.9 },
      { "mode": "generation", "width": 4096, "height": 4096, "references": 0, "seconds": 33.4 },
      { "mode": "edit", "width": 4096, "height": 4096, "references": 1, "seconds": 39.9 },
      { "mode": "edit", "width": 4736, "height": 3552, "references": 1, "seconds": 50.1 },
      { "mode": "edit", "width": 4096, "height": 4096, "references": 3, "seconds": 88.5 }
    ]
  }
}