{
  "schema": "anvil-serving-public-benchmark-summary/v1",
  "date": "2026-08-21",
  "decision": "human-approved-promoted-live-primary",
  "model": {
    "id": "deepseek-ai/DeepSeek-V4-Flash-0731",
    "revision": "9e165c30e2704aec5d9d593cce3eebd58bbef1cb",
    "served_name": "deepseek-v4-flash-0731-infernal-r18-b12x-dspark5-maxseq8-batch4096-tp2-1m"
  },
  "source": {
    "rtx6kpro_revision": "12272507e64752af7a6cde28540edf8ee6c20e59",
    "blackwell_llm_docker_revision": "203e1a11a373a0486897a5a284409ba4417d0f7a",
    "runtime_image_digest": "sha256:414ec7d0d28358cfd8af0697f330f5c8acbb80e4dc4e5ba69c9fd5b5855ea804",
    "vllm_revision": "f0fa1cefc1865d316c2478525f550e7646addc40",
    "b12x_revision": "75787c7a7431b3bea414d2ebf5f2b8671b23eb33",
    "lmcache_revision": "e045d729bc5c4c63a40e13d032f42923de97812f"
  },
  "configuration": {
    "hardware": "2x NVIDIA RTX PRO 6000 Blackwell Max-Q Workstation Edition",
    "topology": "exclusive TP=2 over PCIe without NVLink under WSL2",
    "context_window": 1048576,
    "max_num_seqs": 8,
    "max_num_batched_tokens": 4096,
    "gpu_memory_utilization": 0.975,
    "kv_cache_capacity_tokens": 1323176,
    "full_window_concurrency": 1.26,
    "speculative_method": "fixed probabilistic DSpark K5",
    "native_kv_offload": false,
    "lmcache": false
  },
  "gates": {
    "direct_functional": "pass",
    "routed_functional": {
      "result": "pass",
      "checks": "smoke,json,needle-32k,tools-20-of-20,streaming-tools,tool-result,responses",
      "reasoning_evidence": "any-because-router-does-not-export-upstream-hidden-reasoning"
    },
    "quality": {
      "intelligence": "6/6",
      "session": "3/3",
      "tools": "3/3",
      "failures": 0
    },
    "structured_tool_soak": "160/160",
    "post_reload": "pass",
    "mini_clients": {
      "generation": 2,
      "hermes": "pass-terminal-sentinel-exit-zero-no-raw-dsml",
      "pi": "pass-normal-extension-loaded-process",
      "openclaw": "pass-running-gateway-no-fallback"
    }
  },
  "context": [
    {"actual_prompt_tokens": 354010, "result": "pass", "seconds": 60.882},
    {"actual_prompt_tokens": 585118, "result": "pass", "seconds": 92.087},
    {"actual_prompt_tokens": 900118, "result": "pass", "seconds": 175.999},
    {"actual_prompt_tokens": 1040063, "result": "pass", "seconds": 187.347}
  ],
  "performance": {
    "k5": {
      "4k_c1_decode_p50_tps": 142.075,
      "4k_c1_e2e_p50_ms": 1576,
      "32k_c1_decode_p50_tps": 129.475,
      "32k_c1_e2e_p50_ms": 4098,
      "4k_c8_completed": "8/8",
      "4k_c8_aggregate_output_tps": 272.33,
      "long_c2_completed": "2/2",
      "long_c2_prompt_tokens_per_request": 490861
    },
    "no_spec_control": {
      "4k_c1_decode_p50_tps": 76.4,
      "4k_c1_e2e_p50_ms": 2350,
      "32k_c1_decode_p50_tps": 76.3,
      "32k_c1_e2e_p50_ms": 4660
    }
  },
  "promotion": {
    "state": "live",
    "alias": "llm.primary",
    "context_window": 1048576,
    "mode": "exclusive-tp2",
    "identity": "exact-expected-observed-ready",
    "rollback": {
      "target": "r33-dspark-k5-393216-fixed-port",
      "managed_mode_transition_available": true,
      "automatic_promotion_rollback": "open-p1-not-live-proven"
    }
  },
  "publication_safety": {
    "contains_prompts": false,
    "contains_response_bodies": false,
    "contains_private_topology": false,
    "contains_credentials": false
  }
}
