{
  "schema_version": "1",
  "model": {
    "id": "mlx-community/GLM-4.5-Air-4bit",
    "revision": "60837794f3caafc4682dd1a9188a82c55a9100ef",
    "architecture": "glm4_moe",
    "quantization": {
      "group_size": 64,
      "bits": 4
    },
    "n_experts": 128,
    "top_k": 8,
    "moe_layers": 45
  },
  "strata": {
    "commit": "6562877",
    "mlx_lm": "0.31.3"
  },
  "mode": "streamed",
  "measured_on": {
    "chip": "Apple M5 Max",
    "ram_gb": 137.4,
    "os": "macOS 26.5.1",
    "mlx_lm": "0.31.3"
  },
  "workload": {
    "prompt_set": "bench/mixed-v1",
    "prompt": "Explain how a CPU pipeline works.",
    "generated_tokens": 140,
    "warmup_tokens": 40,
    "decoding": "greedy"
  },
  "memory": {
    "target_ram_gb": 16.0,
    "reserve_gb": 4.0,
    "available_gb": 12.0,
    "enforced_memory_gb": null,
    "full_footprint_gb": 60.14,
    "measured_peak_gb": 11.27,
    "peak_rss_gb": 6.77,
    "swap_growth_gb": 0.0,
    "compression_events": 0,
    "expert_budget": 11,
    "working_set_gb": 8.9,
    "mode": "paged"
  },
  "throughput": {
    "tokens_per_s": 2.91,
    "p50_ms": 342.9,
    "p95_ms": 456.2,
    "ssd_reads_mb_per_tok": 2006.0
  },
  "fidelity": {
    "method": "decode-mode teacher-forced, paged vs resident",
    "prompt_set": "arch_fidelity/mixed-v1",
    "budget": 8,
    "positions": 98,
    "agreement_pct": 95.918,
    "perplexity_delta_pct": -0.3986,
    "high_confidence_flips": 0,
    "max_abs_dlogit": 8.0234375,
    "mean_kl": 0.0058584013022482395,
    "bit_exact": false
  },
  "honesty": {
    "device_real": false,
    "physical_constraint_validated": false,
    "measurement_type": "development-host benchmark",
    "memory_peak_scope": "load-through-generation",
    "throughput_is_optimistic": true
  },
  "sample_output": "Of course! Let's break down how a CPU pipeline works using a simple analogy and then dive into the technical details.\n\n### The Core Idea: An Assembly Line\n\nImagine you want to build a toy car. If you had one person do the entire job\u2014get the chassis, put on the wheels, add the seat, and paint it\u2014it would be very slow. That person would have to finish one car completely before starting the next.\n\nNow, imagine an assembly line:\n*   **Station 1:** Gets the chassis.\n*   **Station 2:** Installs the wheels.\n*   **Station 3:** Adds the seat.\n*   **Station 4:** Paints the car"
}