{
  "base_url": "http://127.0.0.1:39021/v1",
  "completed": 6,
  "concurrency": 2,
  "context_distribution": {
    "2048": 6
  },
  "context_seed": 0,
  "context_tokens": 2048,
  "engine": null,
  "failed": 0,
  "failures": [],
  "gpu": null,
  "manifest": null,
  "max_context_tokens": 65536,
  "max_tokens": 128,
  "measurement_protocol": "capacity-v2",
  "metrics": {
    "content_chunks": 289,
    "content_chunks_s": null,
    "e2e_p50_ms": 422.5768000001153,
    "e2e_p95_ms": 515.3733000001921,
    "output_token_sources": {
      "usage": 6
    },
    "output_tokens": 301,
    "prefix_cache_hit_avg": null,
    "throughput_tok_s": 224.0822121157763,
    "ttft_p50_ms": 122.1399000000929,
    "ttft_p95_ms": 164.37170000017431,
    "ttft_samples": 6
  },
  "model": "nemotron3-omni-30b-a3b-nvfp4",
  "requests": 6,
  "run_id": "benchmark-20260728T025918Z",
  "schema": "anvil-serving.benchmark/v1",
  "serve_flags": {
    "no_thinking": true,
    "reasoning_effort": null,
    "shared_prefix_burst": false,
    "thinking_mode": "disabled"
  },
  "source_recipe": null,
  "tier": null
}