{
  "model": "Sarvam 30B Q4_K_M",
  "hardware": {
    "chip": "Apple M2",
    "unifiedMemoryGiB": 8,
    "power": "AC"
  },
  "date": "2026-10-25",
  "profile": {
    "activeExperts": 3,
    "approximateExecution": true,
    "contextLimitTokens": 2048,
    "expertSlotsPerLayer": 12,
    "ssdReadWorkers": 2,
    "expertStorage": "Verified internal SSD mirror",
    "kvPrecision": "FP32"
  },
  "workload": "Same short Python is_even request, two natural completions, KV cleared between turns",
  "turns": [
    {
      "decodeSteps": 347,
      "decodeSeconds": 82.66584837499977,
      "tokensPerSecond": 4.197622196120127,
      "outputTokens": 348,
      "finish": "stop",
      "peakSampledEngineFootprintBytes": 1915750632
    },
    {
      "decodeSteps": 347,
      "decodeSeconds": 80.73468374999356,
      "tokensPerSecond": 4.298028850580933,
      "outputTokens": 348,
      "finish": "stop",
      "peakSampledEngineFootprintBytes": 1922091264
    }
  ],
  "weightedDecodeTokensPerSecond": 4.247232190585061,
  "qualifiedAcrossWorkloads": false,
  "limitations": [
    "Experimental reduced-expert execution can change answer quality",
    "Decode includes reasoning and excludes prefill and startup",
    "Sampled engine footprint excludes total system and file-cache memory",
    "Not a benchmark for Nemotron, GPT-OSS, Qwen or other models",
    "Separate mixed-language completions were slower and ended during reasoning"
  ],
  "sourceDocument": "https://github.com/MathObsession/Particle/blob/main/docs/PARTICLE_SPEED_QUALITY.md"
}