{
  "schema": "fjson-public-miniv-flash-next/v1",
  "published_at": "2026-09-14",
  "measurement_date": "2026-09-12",
  "hardware": {
    "name": "Corsair AI Workstation 300",
    "memory": "64GB LPDDR5X UMA",
    "gpu": "Radeon 8050S · gfx1151",
    "backend": "Vulkan RADV",
    "gtt": "62 GiB configured"
  },
  "model": {
    "alias": "qwen38-flash-next-fast",
    "name": "Qwen3.8 Flash-Next UD-IQ3_XXS",
    "quant": "UD-IQ3_XXS + MTP Q8_0",
    "model_size_gib": 76.1,
    "status": "64K on-demand validated; 128K benchmark-only; non-default",
    "protected_default": "architect-35b-q6"
  },
  "recipe": {
    "context_tokens": 131072,
    "moe_cache_slots": 48,
    "kv_cache": "q8_0 / q8_0",
    "parallel": 1,
    "draft_n_max": 2
  },
  "benchmark": {
    "measurement_grain": "three fresh-load repetitions per context/cache rung",
    "baseline": {
      "context_tokens": 32768,
      "moe_cache_slots": 0,
      "decode_tokens_per_second": 1.57
    },
    "rungs": [
      {
        "context_tokens": 32768,
        "moe_cache_slots": 32,
        "decode_tokens_per_second": 32.27,
        "exact_content": "3/3"
      },
      {
        "context_tokens": 65536,
        "moe_cache_slots": 32,
        "decode_tokens_per_second": 34.84,
        "exact_content": "3/3"
      },
      {
        "context_tokens": 131072,
        "moe_cache_slots": 32,
        "decode_tokens_per_second": 35.76,
        "exact_content": "3/3"
      },
      {
        "context_tokens": 131072,
        "moe_cache_slots": 48,
        "decode_tokens_per_second_range": [
          36.33,
          37.92
        ],
        "exact_content": "3/3"
      }
    ],
    "prefill_samples": [
      {
        "prompt_tokens": 5660,
        "seconds": 37.1,
        "prompt_tokens_per_second_approx": 164
      },
      {
        "prompt_tokens": 16860,
        "seconds": 76.0,
        "prompt_tokens_per_second_approx": 232
      }
    ],
    "mtp_acceptance_range": [
      0.78,
      0.79
    ],
    "runs_without_device_lost": 13,
    "headline_speedup_vs_baseline": "24.2x decode at the best measured 128K/48-slot rung"
  },
  "scope": [
    "Benchmark throughput and exact-content results are dated observations, not universal quality scores.",
    "The bounded application sample is reported separately and is not averaged with the fresh-load benchmark rungs.",
    "64K is the on-demand operational ceiling; 128K remains benchmark-only. Architect remains the protected default.",
    "Prefill and decode are reported separately."
  ],
  "credit": "All benchmark orchestration, methodology design, and AI model evaluations directed by N30",
  "integration_sample": {
    "result": "PASS",
    "scope": "one bounded application run",
    "completion_tokens": 380,
    "content_words": 309,
    "decode_tokens_per_second": 32.48115578750777,
    "prompt_tokens_per_second": 33.81148407056457,
    "wall_seconds": 13.6,
    "finish_reason": "stop",
    "protected_default_restored": true
  }
}