{
  "schema": "benchmark-summary/v1",
  "observed_date": "2026-08-14",
  "hardware": {
    "gpu": "NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition",
    "count": 2,
    "memory_gib_each": 96,
    "topology": "split TP=1; one independent serve per equal PCIe card"
  },
  "runtime": {
    "image_digest": "sha256:4a2f33a884222f7049b983263ad9976f89452bb81affecf5b67d89ad35c1bc31",
    "vllm_revision": "3a0914114705fa38d4c3171d0746c1a6b6f10209"
  },
  "profiles": [
    {
      "capability": "primary text",
      "model": "Qwen/Qwen3.8-27B-FP8",
      "revision": "017b9c7af6b5689d5dd426a76e0bc077eb5ca20a",
      "weights": "official FP8",
      "kv_cache": "FP8",
      "context_tokens": 393216,
      "tensor_parallel": 1,
      "mtp_speculative_tokens": 3,
      "max_sequences": 1,
      "batch_tokens": 4096,
      "four_k": {
        "requests": 10,
        "median_ttft_seconds": 0.834,
        "median_e2e_seconds": 1.295,
        "decode_tokens_per_second": 93.6
      },
      "routed_functional": {
        "short_coding": "pass",
        "structured_json": "pass",
        "tools": "20/20",
        "streaming_tool": "pass",
        "tool_result_recovery": "pass",
        "responses_subset": "pass"
      }
    },
    {
      "capability": "multimodal and OCR",
      "model": "Qwen/Qwen3.8-27B",
      "revision": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
      "weights": "BF16",
      "kv_cache": "FP8",
      "context_tokens": 393216,
      "tensor_parallel": 1,
      "mtp_speculative_tokens": 3,
      "max_sequences": 1,
      "batch_tokens": 4096,
      "images_per_request": 32,
      "four_k": {
        "requests": 10,
        "median_ttft_seconds": 0.884,
        "median_e2e_seconds": 1.584,
        "decode_tokens_per_second": 62.0
      },
      "routed_multimodal": {
        "passed": 30,
        "attempts": 30,
        "image": {
          "passed": 12,
          "attempts": 12,
          "median_seconds": 0.9668275
        },
        "video": {
          "passed": 14,
          "attempts": 14,
          "median_seconds": 2.218448
        },
        "mixed": {
          "passed": 4,
          "attempts": 4,
          "median_seconds": 2.621022
        }
      },
      "thirty_two_image_request": {
        "passed": true,
        "seconds": 5.71659,
        "prompt_tokens": 7173,
        "completion_tokens": 265
      },
      "routed_ocr": "pass"
    }
  ],
  "client_acceptance": {
    "hermes": {
      "text": "pass",
      "image": "pass",
      "fallback_used": false
    },
    "openclaw": {
      "primary": "pass",
      "vision_alias": "pass",
      "fallback_used": false,
      "reported_context_tokens": 393216
    }
  },
  "retained_failures": [
    "The first routed multimodal run passed 22/30 at a 512-token cap because the router did not preserve the caller's chat_template_kwargs extension.",
    "A 2048-token retry passed 28/30 but still exhausted two verbose video attempts while thinking remained enabled upstream.",
    "After installing a thinking-disabled soft default, the unchanged 512-token corpus passed 30/30.",
    "The Responses subset rejects chat-only chat_template_kwargs; its separate supported-field request passed."
  ]
}