{
  "title": "Inside a Fast Local AI Workstation",
  "article_date": "2026-09-29",
  "benchmark_dates": [
    "2026-09-13",
    "2026-09-25"
  ],
  "note": "Selected local records reviewed for the article; inference benchmarks were not rerun on publication day. Source fingerprints refer to original private files, which are not included. No raw conversations are published.",
  "hardware": {
    "gpu": "NVIDIA GeForce RTX 5090",
    "gpu_reported_mib": 32607,
    "cpu": "AMD Ryzen 9950X3D",
    "system_ram_gb": 96
  },
  "runtime": {
    "llama_cpp": "b11146",
    "cuda": "13.4",
    "model": "Qwen3.8-27B Q4_K_M"
  },
  "settings": {
    "context": 524288,
    "cache_k": "q4_0",
    "cache_v": "q4_0",
    "draft_cache_k": "q4_0",
    "draft_cache_v": "q4_0",
    "batch": 2048,
    "ubatch": 128,
    "threads": 16,
    "speculation": "draft-mtp",
    "draft_tokens": 3,
    "rope_scale": 2,
    "yarn_original_context": 262144,
    "cache_ram_mib": 16384,
    "context_checkpoints": 8,
    "checkpoint_min_tokens": 65536,
    "projector_gpu": false
  },
  "speed": {
    "initial": {
      "workloads": [
        {
          "lines": 64,
          "samples": 3,
          "median_decode_tps": 34.42108454096967,
          "median_first_content_s": 1.7406946999253705,
          "median_wall_s": 12.874209300032817,
          "median_prefill_tps": 976.6009931914439
        },
        {
          "lines": 512,
          "samples": 3,
          "median_decode_tps": 31.534728546092573,
          "median_first_content_s": 18.595771799911745,
          "median_wall_s": 30.808830599999055,
          "median_prefill_tps": 714.0106055051734
        }
      ],
      "passed": true,
      "concurrent_ollama_detected": false,
      "gpu_peak_mib": 32058
    },
    "microbatch128": {
      "workloads": [
        {
          "lines": 64,
          "samples": 3,
          "median_decode_tps": 106.37142624641344,
          "median_first_content_s": 1.286570299998857,
          "median_wall_s": 4.8182052000192925,
          "median_prefill_tps": 1342.7040365796124
        },
        {
          "lines": 512,
          "samples": 3,
          "median_decode_tps": 123.49408953418224,
          "median_first_content_s": 10.425469300011173,
          "median_wall_s": 13.649052100023255,
          "median_prefill_tps": 1276.9903845819654
        }
      ],
      "passed": true,
      "concurrent_ollama_detected": false,
      "gpu_peak_mib": 32068
    },
    "alternative512": {
      "workloads": [
        {
          "lines": 64,
          "samples": 3,
          "median_decode_tps": 86.5613797025091,
          "median_first_content_s": 1.287687600008212,
          "median_wall_s": 5.687619799980894,
          "median_prefill_tps": 1332.3589973537312
        },
        {
          "lines": 512,
          "samples": 3,
          "median_decode_tps": 71.21057040038747,
          "median_first_content_s": 12.30503020004835,
          "median_wall_s": 17.54469289991539,
          "median_prefill_tps": 1078.873596259187
        }
      ],
      "passed": true,
      "concurrent_ollama_detected": false,
      "gpu_peak_mib": 32120
    },
    "boost": {
      "workloads": [
        {
          "lines": 64,
          "samples": 3,
          "median_decode_tps": 143.06039427145842,
          "median_first_content_s": 0.8578497999114916,
          "median_wall_s": 3.535243499907665,
          "median_prefill_tps": 2033.9672531272247
        },
        {
          "lines": 512,
          "samples": 3,
          "median_decode_tps": 137.28233312579684,
          "median_first_content_s": 6.219459099927917,
          "median_wall_s": 9.01046579994727,
          "median_prefill_tps": 2142.532418577209
        }
      ],
      "passed": true,
      "concurrent_ollama_detected": false,
      "gpu_peak_mib": 31421
    },
    "standard": {
      "workloads": [
        {
          "lines": 64,
          "samples": 3,
          "median_decode_tps": 80.48448719298479,
          "median_first_content_s": 0.7953468000050634,
          "median_wall_s": 5.554965500021353,
          "median_prefill_tps": 2241.521596017227
        },
        {
          "lines": 512,
          "samples": 3,
          "median_decode_tps": 75.99857765586256,
          "median_first_content_s": 5.8657865999266505,
          "median_wall_s": 10.914652799954638,
          "median_prefill_tps": 2277.8400036503294
        }
      ],
      "passed": true,
      "concurrent_ollama_detected": false,
      "gpu_peak_mib": 27889
    }
  },
  "retrieval": {
    "input_tokens": 497969,
    "retrieval_passed": true,
    "followup_passed": true,
    "code_followup_tokens": 384,
    "code_followup_decode_tps": 39.289864928009784,
    "passed": true,
    "note": "One seeded retrieval, its cached repeat, and a 384-token code speed probe. Code is deliberately token-limited and not graded for correctness. Not a general long-context quality evaluation.",
    "concurrent_ollama_detected": false,
    "gpu_peak_mib": 32042
  },
  "audit": {
    "passed": true,
    "server_input_tokens": 497969,
    "cold_first_content_s": 611.3449246999808,
    "cold_total_s": 612.3728844999569,
    "cached_first_content_s": 0.8030258000362664,
    "cached_total_s": 1.8166557000949979,
    "cached_fraction": 0.9999498034302328,
    "large_history_code_tps": 39.289864928009784,
    "verification": "All server input counts match tokenization and cache+prefill totals. Both retrieval answers match expected values, stop at EOS and are not truncated. Code benchmark has its intended 384-token limit."
  },
  "wrapper": {
    "kind": "same-engine-wrapper-overhead",
    "concurrent_ollama_detected": false,
    "samples": 20,
    "results": [
      {
        "lines": 64,
        "path": "ollama",
        "samples": 5,
        "decode_tps": {
          "median": 178.1581784810902,
          "min": 172.83626104756777,
          "max": 183.03756568762424
        },
        "wall_s": {
          "median": 2.8126778999576345,
          "min": 2.753864899976179,
          "max": 2.8832770000444725
        },
        "first_content_s": {
          "median": 0.6570678999414667,
          "min": 0.6408571000210941,
          "max": 0.6609649000456557
        },
        "prefill_s": {
          "median": 0.595241,
          "min": 0.590709,
          "max": 0.610889
        },
        "prompt_tokens": {
          "median": 1656,
          "min": 1652,
          "max": 1661
        }
      },
      {
        "lines": 64,
        "path": "direct",
        "samples": 5,
        "decode_tps": {
          "median": 182.30095133561554,
          "min": 180.80389933742907,
          "max": 183.35280661101467
        },
        "wall_s": {
          "median": 2.7440531000029296,
          "min": 2.735125899896957,
          "max": 2.783070100005716
        },
        "first_content_s": {
          "median": 0.6405287999659777,
          "min": 0.6351456999545917,
          "max": 0.6589136999100447
        },
        "prefill_s": {
          "median": 0.596751,
          "min": 0.587352,
          "max": 0.6063569999999999
        },
        "prompt_tokens": {
          "median": 1656,
          "min": 1652,
          "max": 1661
        }
      },
      {
        "lines": 512,
        "path": "ollama",
        "samples": 5,
        "decode_tps": {
          "median": 159.69969802616995,
          "min": 158.40447096619303,
          "max": 162.66251425415
        },
        "wall_s": {
          "median": 6.841986700077541,
          "min": 6.684037800063379,
          "max": 6.877289200085215
        },
        "first_content_s": {
          "median": 4.437003600061871,
          "min": 4.322674600058235,
          "max": 4.45171710010618
        },
        "prefill_s": {
          "median": 4.378175,
          "min": 4.275237,
          "max": 4.4036
        },
        "prompt_tokens": {
          "median": 13222,
          "min": 13216,
          "max": 13224
        }
      },
      {
        "lines": 512,
        "path": "direct",
        "samples": 5,
        "decode_tps": {
          "median": 160.2660416291043,
          "min": 158.90343388665386,
          "max": 164.72231421124607
        },
        "wall_s": {
          "median": 6.835859999991953,
          "min": 6.711020800052211,
          "max": 6.83885340008419
        },
        "first_content_s": {
          "median": 4.421453200047836,
          "min": 4.375002700020559,
          "max": 4.439372200053185
        },
        "prefill_s": {
          "median": 4.37257,
          "min": 4.3172619999999995,
          "max": 4.385185000000001
        },
        "prompt_tokens": {
          "median": 13222,
          "min": 13216,
          "max": 13224
        }
      }
    ],
    "output_equal_pairs": 3,
    "runner_launch_count": 1
  },
  "integration": {
    "routing": {
      "kind": "router-mock",
      "passes": 5,
      "attempts": 5
    },
    "reasoning": {
      "kind": "stack-probe",
      "passes": 5,
      "attempts": 5
    },
    "nativeSmoke": {
      "tool_passed": true,
      "vision_passed": true,
      "passed": true
    },
    "javascriptRepair": {
      "tests_unchanged": true,
      "model": "coding",
      "passed": true,
      "agent_exit": 0,
      "test_exit": 0,
      "fixture": "absolute TTL repair; one file and five assertions",
      "elapsed_s": 7.3503756
    },
    "rustRepair": {
      "elapsed_s": 7.6086401,
      "agent_exit": 0,
      "tests_unchanged": true,
      "fixture": "Rust parser error priority; five test groups; based on observed holdout failure",
      "model": "coding",
      "compile_exit": 0,
      "passed": true,
      "test_exit": 0
    }
  },
  "historicalQuality": {
    "core": {
      "kind": "quality",
      "attempts": 9,
      "tasks_passed": 9,
      "assertions_passed": 47,
      "assertions_total": 47,
      "errors": 0,
      "truncated": 0
    },
    "holdout": {
      "kind": "quality",
      "attempts": 3,
      "tasks_passed": 2,
      "assertions_passed": 16,
      "assertions_total": 17,
      "errors": 0,
      "truncated": 0
    },
    "strictTools": {
      "kind": "tools",
      "attempts": 2,
      "passes": 0,
      "errors": 0,
      "note": "Two-turn synthetic file read, argument correctness, tool-result handling, and one injection fixture. No real filesystem tools are executed by this suite."
    }
  },
  "sourceRecords": [
    {
      "file": "results/llamacpp-migration-20260925/speed-clean-u512-mtp3/summary.json",
      "sha256": "5e70e7c8471f9bfdbdcaa205eac248da11648c154d56f408d4708b4eb71298af"
    },
    {
      "file": "results/llamacpp-migration-20260925/speed-clean-u128-mtp3/summary.json",
      "sha256": "10c597c44640a040210e0690ba0908ea3b75fcbdc567169aafe72cb2a5292d30"
    },
    {
      "file": "results/llamacpp-migration-20260925/speed-clean-b512-mtp3/summary.json",
      "sha256": "953443d5cf2dafb768f87842ae78c959daf1a80b4146ac5c01d7e15c3bdf07bd"
    },
    {
      "file": "results/llamacpp-migration-20260925/speed-clean-headroom-mtp3/summary.json",
      "sha256": "b012c26d34a4302a9b5a0de786de5fdcb802b17139a7a399167640d70f6a4a08"
    },
    {
      "file": "results/llamacpp-migration-20260925/speed-clean-headroom-standard/summary.json",
      "sha256": "1657bf37f872918414c7a7be5672e1eb6cb964948a73dce6031932408ccdf4f4"
    },
    {
      "file": "results/llamacpp-migration-20260925/context-500k/summary.json",
      "sha256": "7ad1fcc5585dbcb26518977676e13d5d336e4ffb9a10b86dc0af8672930b78f1"
    },
    {
      "file": "results/llamacpp-migration-20260925/context-500k/audit.json",
      "sha256": "8ae6f5466ca47a05f36bf1b5e4b69274429ef97333b3660a31e84f9f116fb433"
    },
    {
      "file": "results/wrapper-overhead-20260925/controlled/summary.json",
      "sha256": "655ec10629469402cddb72ea5e65579f606dd188ca5a9178baeeeee5efea3b5c"
    },
    {
      "file": "results/llama512-js/summary.json",
      "sha256": "20ad4bb941b89e80e2919acecc7462f13fb68923fe859b056a391c282d333cc1"
    },
    {
      "file": "results/llama512-rust/summary.json",
      "sha256": "c2b98ff8668bf9d2c320c0d8660d4dbbb4df39806619601167f2b821f9845fec"
    },
    {
      "file": "results/llamacpp-migration-20260925/routing.json",
      "sha256": "beb7709bddf0790a872bd5ff989d1b5c4da19a09581aafa8b4e62e584e69c481"
    },
    {
      "file": "results/llamacpp-migration-20260925/stack-final.json",
      "sha256": "45b24356444d8d00656526765b09743bd8fdf152df63eb66d1299e758e91edbf"
    },
    {
      "file": "results/llamacpp-migration-20260925/native-smoke/summary.json",
      "sha256": "befaf1486e9a1f4ede2dc5df49566fc11eaf5ab2ca85e7c3448310c0b47facf4"
    },
    {
      "file": "results/01-baseline-default-context/summary.json",
      "sha256": "145083848bbc0172f8b52428f3569f912f9e0c75769b0c9b8501e459e25fdacb"
    },
    {
      "file": "results/19-standard-unreal-closed/summary.json",
      "sha256": "bc01b96bf4f1ed9cfa4217872424ee76bb24c1537e0f5340dba5619d0f42265d"
    },
    {
      "file": "results/20-boost-unreal-closed/summary.json",
      "sha256": "ff4550bc7a3ed0225fe6f0ba0f79fce10a7615d4519edab281f9768fc516ba19"
    },
    {
      "file": "results/01-baseline-default-context/requests.jsonl",
      "sha256": "53a134d1440b82ee4ac50dd97e1207b063b84d63312dbf8c484ffa1bcb548e14"
    },
    {
      "file": "results/08-mtp-greedy/summary.json",
      "sha256": "f6fae68a8046e71fb2f80ab01bd3cc250d457fa782ad3a7b45f76491465f4334"
    },
    {
      "file": "results/09-mtp-greedy-holdout/summary.json",
      "sha256": "d7907c8e244f0cdd7e33c5015cde0f6a9dbab55a74b4cb20739e27d7cc546a03"
    },
    {
      "file": "results/14-no-mtp-tools/summary.json",
      "sha256": "32f6d6239790f16aec2d190f2acb053b3ef91227f7f56b4a52950c5e9ce7478e"
    },
    {
      "file": "config/llama.json",
      "sha256": "51ab29a3f953d0c7dfeba075433494db8f8aa020fc9be735eb575556ae6b36b1"
    }
  ],
  "historicalSpeed": {
    "original": {
      "kind": "speed",
      "model": "qwen3.8:27b-q4_K_M",
      "options": {
        "seed": 42,
        "num_predict": 384,
        "temperature": 0.2
      },
      "think": "false",
      "warm_samples": 3,
      "errors": 0,
      "decode_tps": {
        "median": 77.64301681941852,
        "min": 77.47113242015581,
        "max": 77.94086723828715
      },
      "wall_s": {
        "median": 5.581580199999735,
        "min": 5.546694500022568,
        "max": 5.597923699999228
      },
      "first_content_s": {
        "median": 0.6359180000144988,
        "min": 0.6199905999819748,
        "max": 0.6413025000365451
      },
      "requestAudit": [
        {
          "case": "cold",
          "allocated_context": 131072,
          "input_tokens": 54,
          "cached_tokens": 0,
          "output_tokens": 384
        },
        {
          "case": "warm-1",
          "allocated_context": 131072,
          "input_tokens": 1643,
          "cached_tokens": 0,
          "output_tokens": 384
        },
        {
          "case": "warm-2",
          "allocated_context": 131072,
          "input_tokens": 1640,
          "cached_tokens": 0,
          "output_tokens": 384
        },
        {
          "case": "warm-3",
          "allocated_context": 131072,
          "input_tokens": 1642,
          "cached_tokens": 0,
          "output_tokens": 384
        }
      ]
    },
    "matchedStandard64k": {
      "kind": "speed",
      "model": "local-coder:fast",
      "options": {
        "seed": 42,
        "num_predict": 384,
        "num_ctx": 65536,
        "num_batch": 512,
        "temperature": 0,
        "top_p": 0.95,
        "top_k": 20,
        "min_p": 0,
        "repeat_penalty": 1,
        "presence_penalty": 0
      },
      "think": "false",
      "warm_samples": 5,
      "errors": 0,
      "decode_tps": {
        "median": 79.49426740753668,
        "min": 79.3417937030716,
        "max": 79.861839018498
      },
      "wall_s": {
        "median": 5.463653399958275,
        "min": 5.41549710003892,
        "max": 5.476675499987323
      },
      "first_content_s": {
        "median": 0.6275801000301726,
        "min": 0.6068440999952145,
        "max": 0.6366177999880165
      }
    },
    "optimized64k": {
      "kind": "speed",
      "model": "local-coder:fast",
      "options": {
        "seed": 42,
        "num_predict": 384,
        "num_ctx": 65536,
        "num_batch": 512,
        "temperature": 0,
        "top_p": 0.95,
        "top_k": 20,
        "min_p": 0,
        "repeat_penalty": 1,
        "presence_penalty": 0
      },
      "think": "false",
      "warm_samples": 5,
      "errors": 0,
      "decode_tps": {
        "median": 149.08481722550823,
        "min": 131.84712326057937,
        "max": 153.81799085664204
      },
      "wall_s": {
        "median": 3.2601977000012994,
        "min": 3.184532899991609,
        "max": 3.5976442999672145
      },
      "first_content_s": {
        "median": 0.6777161000063643,
        "min": 0.6347606999916025,
        "max": 0.6842416999861598
      }
    }
  },
  "originalToCurrent": {
    "comparison": "Historical endpoints, not an isolated causal experiment",
    "decode_improvement_pct": 84.25403871694874,
    "wall_reduction_pct": 36.662318317887234,
    "first_content_increase_pct": 34.89943670283475,
    "context_allocation_ratio": 4
  }
}