{
  "id": "profiles/turboquant",
  "name": "turboquant",
  "description": "Experimental tiered TurboQuant profile with execution-v1-native session KV cache settings.",
  "intent": "calibrate",
  "compatibleIntents": ["calibrate"],
  "stability": "experimental",
  "owner": "inference",
  "createdAtUtc": "2026-03-25T00:00:00Z",
  "extends": "profiles/throughput",
  "runtime": {
    "inference": {
      "session": {
        "kvcache": {
          "layout": "tiered",
          "maxSeqLen": 2048,
          "kvDtype": "f16",
          "tiering": {
            "mode": "turboquant",
            "hotWindow": 256,
            "coldPageSize": 64,
            "coldDtype": "f16",
            "compression": {
              "mode": "turboquant",
              "blockSize": 1,
              "bitWidth": 4,
              "prodMode": false
            },
            "gating": {
              "mode": "force_on"
            }
          }
        }
      }
    }
  }
}