{
  "schema": "modelfit.local-bench-collection.v1",
  "created_at": "2026-11-21T16:44:35.194731+00:00",
  "model": "Inkling 2 Mini",
  "model_repo": "netprojects/Inkling-2-Mini-MLX-4-8bit",
  "revision": "67a7c329731b55ef3dc116ac697753201645bcd3",
  "machine": {
    "model": "Mac17,16",
    "chip": "Apple M5 Pro",
    "ram_gb": 48,
    "arch": "arm64"
  },
  "macos": "27.2.1",
  "software": {
    "python": "3.13.18",
    "mlx": "0.32.4",
    "mlx_vlm": "0.8.0",
    "mlx_vlm_source": "https://github.com/Neywa-Labs/mlx-vlm",
    "mlx_vlm_commit": "434a1bedd1e8a9ec9588462e112e6a0e58696872"
  },
  "weights_bytes": 19526043144,
  "draft_bytes_not_loaded": 635801416,
  "protocol": {
    "reasoning_effort": "none",
    "sampling": "greedy",
    "temperature": 0.0,
    "max_tokens": 256,
    "runs_per_prompt": 3,
    "modalities": "text only",
    "draft_model": null,
    "weights_downloaded_during_bench": false,
    "offline_env": {
      "HF_HUB_OFFLINE": "1",
      "TRANSFORMERS_OFFLINE": "1",
      "HF_DATASETS_OFFLINE": "1",
      "HF_HUB_DISABLE_TELEMETRY": "1"
    }
  },
  "prompts": {
    "chat": "Give five practical tips for keeping a laptop comfortable during long local AI inference runs. Explain each tip in one or two sentences.",
    "code": "Write a complete Python command-line program that converts between decimal gigabytes and binary gibibytes. Include argument parsing, input validation, two conversion directions, and useful error messages.",
    "reasoning": "A train travels 120 km in 1.5 hours, then 80 km in 45 minutes. Calculate its average speed for the whole journey. Show the calculation, check it with a second method, and explain one common mistake."
  },
  "commands": [
    "/Volumes/MINI2TO/scratch/dev-inkling2/venv/bin/python /Volumes/MINI2TO/scratch/dev-inkling2/bench.py --model /Volumes/MINI2TO/models/huggingface/hub/models--netprojects--Inkling-2-Mini-MLX-4-8bit/snapshots/67a7c329731b55ef3dc116ac697753201645bcd3 --output /Volumes/MINI2TO/scratch/dev-inkling2/bench-default.json --runs 3 --max-tokens 256",
    "/Volumes/MINI2TO/scratch/dev-inkling2/venv/bin/python /Volumes/MINI2TO/scratch/dev-inkling2/bench.py --model /Volumes/MINI2TO/models/huggingface/hub/models--netprojects--Inkling-2-Mini-MLX-4-8bit/snapshots/67a7c329731b55ef3dc116ac697753201645bcd3 --output /Volumes/MINI2TO/scratch/dev-inkling2/bench-prefill256.json --runs 3 --max-tokens 256 --prefill-step-size 256"
  ],
  "mq_jobs": [
    "inkling2-bench-default",
    "inkling2-bench-prefill256"
  ],
  "benchmark_script_sha256": "32ca263bf151e94c5052a0c243c0c9a86640702cb569b2eee2c9e603e9830606",
  "benchmark_script": "#!/usr/bin/env python3\n\"\"\"Reproducible Inkling 2 Mini local MLX bench. No downloads, no draft model.\"\"\"\nimport argparse\nimport dataclasses\nimport datetime as dt\nimport importlib.metadata as metadata\nimport json\nimport os\nimport platform\nimport statistics\nimport subprocess\nimport time\nimport traceback\nfrom pathlib import Path\n\nPROMPTS = {\n    \"chat\": \"Give five practical tips for keeping a laptop comfortable during long local AI inference runs. Explain each tip in one or two sentences.\",\n    \"code\": \"Write a complete Python command-line program that converts between decimal gigabytes and binary gibibytes. Include argument parsing, input validation, two conversion directions, and useful error messages.\",\n    \"reasoning\": \"A train travels 120 km in 1.5 hours, then 80 km in 45 minutes. Calculate its average speed for the whole journey. Show the calculation, check it with a second method, and explain one common mistake.\",\n}\n\n\ndef run(cmd):\n    return subprocess.run(cmd, capture_output=True, text=True).stdout.strip()\n\n\ndef main():\n    ap = argparse.ArgumentParser()\n    ap.add_argument(\"--model\", required=True)\n    ap.add_argument(\"--output\", required=True)\n    ap.add_argument(\"--runs\", type=int, default=3)\n    ap.add_argument(\"--max-tokens\", type=int, default=256)\n    ap.add_argument(\"--prefill-step-size\", type=int, default=None)\n    args = ap.parse_args()\n    out = Path(args.output)\n    out.parent.mkdir(parents=True, exist_ok=True)\n    model_path = Path(args.model)\n    record = {\n        \"schema\": \"modelfit.local-bench.v1\",\n        \"model\": \"Inkling 2 Mini\",\n        \"model_repo\": \"netprojects/Inkling-2-Mini-MLX-4-8bit\",\n        \"revision\": model_path.name,\n        \"model_path\": str(model_path),\n        \"machine\": {\"model\": run([\"sysctl\", \"-n\", \"hw.model\"]), \"chip\": \"Apple M5 Pro\", \"ram_gb\": 48, \"arch\": platform.machine()},\n        \"software\": {\n            \"python\": platform.python_version(),\n            \"mlx\": metadata.version(\"mlx\"),\n            \"mlx_vlm\": metadata.version(\"mlx-vlm\"),\n            \"mlx_vlm_source\": \"https://github.com/Neywa-Labs/mlx-vlm\",\n            \"mlx_vlm_commit\": os.environ.get(\"MLX_VLM_COMMIT\", \"unknown\"),\n        },\n        \"settings\": {\n            \"reasoning_effort\": \"none\",\n            \"temperature\": 0.0,\n            \"sampling\": \"greedy\",\n            \"max_tokens\": args.max_tokens,\n            \"runs_per_prompt\": args.runs,\n            \"prefill_step_size\": args.prefill_step_size,\n            \"draft_model\": None,\n            \"offline\": True,\n        },\n        \"command\": \" \".join([os.environ.get(\"PYTHON\", \"python\"), *os.sys.argv]),\n        \"prompts\": PROMPTS,\n        \"started_at\": dt.datetime.now(dt.timezone.utc).isoformat(),\n        \"status\": \"running\",\n        \"runs\": [],\n    }\n    out.write_text(json.dumps(record, indent=2) + \"\\n\")\n    try:\n        import mlx.core as mx\n        from mlx_vlm import load, generate\n        from mlx_vlm.prompt_utils import apply_chat_template\n\n        weights = sum(p.stat().st_size for p in model_path.glob(\"model-*.safetensors\"))\n        draft = model_path / \"mtp.safetensors\"\n        record[\"weights_bytes\"] = weights\n        record[\"draft_bytes_not_loaded\"] = draft.stat().st_size if draft.exists() else None\n        t0 = time.perf_counter()\n        model, processor = load(str(model_path))\n        mx.eval(mx.array([0]))\n        record[\"load_seconds\"] = time.perf_counter() - t0\n        record[\"memory_after_load_bytes\"] = int(mx.metal.get_active_memory())\n        for name, text in PROMPTS.items():\n            prompt = apply_chat_template(processor, model.config, text, reasoning_effort=\"none\")\n            for i in range(args.runs):\n                mx.metal.reset_peak_memory()\n                kwargs = {\"max_tokens\": args.max_tokens, \"temperature\": 0.0}\n                if args.prefill_step_size is not None:\n                    kwargs[\"prefill_step_size\"] = args.prefill_step_size\n                started = time.perf_counter()\n                result = generate(model, processor, prompt, **kwargs)\n                elapsed = time.perf_counter() - started\n                item = {\n                    \"prompt\": name,\n                    \"run\": i + 1,\n                    \"elapsed_seconds\": elapsed,\n                    \"prompt_tokens\": result.prompt_tokens,\n                    \"generation_tokens\": result.generation_tokens,\n                    \"prompt_tps\": result.prompt_tps,\n                    \"generation_tps\": result.generation_tps,\n                    \"peak_memory_gb_reported\": result.peak_memory,\n                    \"peak_memory_bytes_mx\": int(mx.metal.get_peak_memory()),\n                    \"finish_reason\": result.finish_reason,\n                    \"output_text\": result.text,\n                }\n                record[\"runs\"].append(item)\n                out.write_text(json.dumps(record, indent=2) + \"\\n\")\n                print(json.dumps({k: v for k, v in item.items() if k != \"output_text\"}), flush=True)\n        record[\"status\"] = \"complete\"\n        record[\"finished_at\"] = dt.datetime.now(dt.timezone.utc).isoformat()\n        for name in PROMPTS:\n            subset = [r for r in record[\"runs\"] if r[\"prompt\"] == name]\n            record.setdefault(\"medians\", {})[name] = {\n                \"prompt_tokens\": statistics.median(r[\"prompt_tokens\"] for r in subset),\n                \"generation_tokens\": statistics.median(r[\"generation_tokens\"] for r in subset),\n                \"prompt_tps\": statistics.median(r[\"prompt_tps\"] for r in subset),\n                \"generation_tps\": statistics.median(r[\"generation_tps\"] for r in subset),\n                \"peak_memory_gb_reported\": statistics.median(r[\"peak_memory_gb_reported\"] for r in subset),\n                \"peak_memory_bytes_mx\": statistics.median(r[\"peak_memory_bytes_mx\"] for r in subset),\n            }\n        out.write_text(json.dumps(record, indent=2) + \"\\n\")\n    except Exception as exc:\n        record[\"status\"] = \"error\"\n        record[\"error\"] = f\"{type(exc).__name__}: {exc}\"\n        record[\"traceback\"] = traceback.format_exc()\n        record[\"finished_at\"] = dt.datetime.now(dt.timezone.utc).isoformat()\n        out.write_text(json.dumps(record, indent=2) + \"\\n\")\n        print(record[\"traceback\"], flush=True)\n        raise\n\n\nif __name__ == \"__main__\":\n    main()\n",
  "derived_memory_headroom": {
    "basis": "worst median default-run MLX-reported peak, decimal GB; before macOS and other apps",
    "peak_gb": 21.6,
    "formula": "installed GB - peak GB",
    "by_installed_memory_gb": {
      "32": 10.4,
      "36": 14.4,
      "48": 26.4,
      "64": 42.4,
      "128": 106.4
    }
  },
  "runs": [
    {
      "schema": "modelfit.local-bench.v1",
      "model": "Inkling 2 Mini",
      "model_repo": "netprojects/Inkling-2-Mini-MLX-4-8bit",
      "revision": "67a7c329731b55ef3dc116ac697753201645bcd3",
      "model_path": "/Volumes/MINI2TO/models/huggingface/hub/models--netprojects--Inkling-2-Mini-MLX-4-8bit/snapshots/67a7c329731b55ef3dc116ac697753201645bcd3",
      "machine": {
        "model": "Mac17,16",
        "chip": "Apple M5 Pro",
        "ram_gb": 48,
        "arch": "arm64"
      },
      "software": {
        "python": "3.13.18",
        "mlx": "0.32.4",
        "mlx_vlm": "0.8.0",
        "mlx_vlm_source": "https://github.com/Neywa-Labs/mlx-vlm",
        "mlx_vlm_commit": "434a1bedd1e8a9ec9588462e112e6a0e58696872"
      },
      "settings": {
        "reasoning_effort": "none",
        "temperature": 0.0,
        "sampling": "greedy",
        "max_tokens": 256,
        "runs_per_prompt": 3,
        "prefill_step_size": null,
        "draft_model": null,
        "offline": true
      },
      "command": "/Volumes/MINI2TO/scratch/dev-inkling2/venv/bin/python /Volumes/MINI2TO/scratch/dev-inkling2/bench.py --model /Volumes/MINI2TO/models/huggingface/hub/models--netprojects--Inkling-2-Mini-MLX-4-8bit/snapshots/67a7c329731b55ef3dc116ac697753201645bcd3 --output /Volumes/MINI2TO/scratch/dev-inkling2/bench-default.json --runs 3 --max-tokens 256",
      "prompts": {
        "chat": "Give five practical tips for keeping a laptop comfortable during long local AI inference runs. Explain each tip in one or two sentences.",
        "code": "Write a complete Python command-line program that converts between decimal gigabytes and binary gibibytes. Include argument parsing, input validation, two conversion directions, and useful error messages.",
        "reasoning": "A train travels 120 km in 1.5 hours, then 80 km in 45 minutes. Calculate its average speed for the whole journey. Show the calculation, check it with a second method, and explain one common mistake."
      },
      "started_at": "2026-11-21T16:40:06.597092+00:00",
      "status": "complete",
      "runs": [
        {
          "prompt": "chat",
          "run": 1,
          "elapsed_seconds": 17.58105520800018,
          "prompt_tokens": 38,
          "generation_tokens": 256,
          "prompt_tps": 5.903396049657018,
          "generation_tps": 25.325138985608188,
          "peak_memory_gb_reported": 21.511232101,
          "peak_memory_bytes_mx": 21511232101,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for five practical tips for keeping a laptop comfortable during long local AI inference runs, each explained in one or two sentences. Need answer concise. Need ensure safety? This is benign. Need provide practical tips: cooling/ventilation, power/thermal management, workload/batching, posture/ergonomics, monitoring/thermal throttling, maybe use external cooling/stand, reduce precision/quantization, close background apps, use battery/AC. Need exactly five tips, each 1-2 sentences. Mention comfort includes physical and thermal. Let's craft with five bullets. Ensure no overclaim. Mention local AI inference long runs: laptop gets hot, fans loud, battery drains, wrists/neck strain. Tips:\n1. Elevate and ventilate: laptop stand, clear vents, hard flat surface, avoid soft surfaces.\n2. Use AC and performance profile: plug in, disable aggressive battery saver, but avoid max turbo if hot; monitor temps.\n3. Reduce workload: smaller models/quantized/batched inference, lower resolution/context, close background apps.\n4. Ergonomics: external keyboard/mouse, screen height, breaks,"
        },
        {
          "prompt": "chat",
          "run": 2,
          "elapsed_seconds": 9.734617124999204,
          "prompt_tokens": 38,
          "generation_tokens": 256,
          "prompt_tps": 123.34930671135285,
          "generation_tps": 27.183985580403203,
          "peak_memory_gb_reported": 21.5198535,
          "peak_memory_bytes_mx": 21519853500,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for five practical tips for keeping a laptop comfortable during long local AI inference runs, each explained in one or two sentences. Need answer concise. Need ensure safety? This is benign. Need provide practical tips: cooling/ventilation, power/thermal management, workload/batching, posture/ergonomics, monitoring/thermal throttling, maybe use external cooling/stand, reduce precision/quantization, close background apps, use battery/AC. Need exactly five tips, each 1-2 sentences. Mention comfort includes physical and thermal. Let's craft with five bullets. Ensure no overclaim. Mention local AI inference long runs: laptop gets hot, fans loud, battery drains, wrists/neck strain. Tips:\n1. Elevate and ventilate: laptop stand, clear vents, hard flat surface, avoid soft surfaces.\n2. Use AC and performance profile: plug in, disable aggressive battery saver, but avoid max turbo if hot; monitor temps.\n3. Reduce workload: smaller models/quantized/batched inference, lower resolution/context, close background apps.\n4. Ergonomics: external keyboard/mouse, screen height, breaks,"
        },
        {
          "prompt": "chat",
          "run": 3,
          "elapsed_seconds": 9.82729983399986,
          "prompt_tokens": 38,
          "generation_tokens": 256,
          "prompt_tps": 123.33694612191381,
          "generation_tps": 26.919339132396125,
          "peak_memory_gb_reported": 21.520902076,
          "peak_memory_bytes_mx": 21520902076,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for five practical tips for keeping a laptop comfortable during long local AI inference runs, each explained in one or two sentences. Need answer concise. Need ensure safety? This is benign. Need provide practical tips: cooling/ventilation, power/thermal management, workload/batching, posture/ergonomics, monitoring/thermal throttling, maybe use external cooling/stand, reduce precision/quantization, close background apps, use battery/AC. Need exactly five tips, each 1-2 sentences. Mention comfort includes physical and thermal. Let's craft with five bullets. Ensure no overclaim. Mention local AI inference long runs: laptop gets hot, fans loud, battery drains, wrists/neck strain. Tips:\n1. Elevate and ventilate: laptop stand, clear vents, hard flat surface, avoid soft surfaces.\n2. Use AC and performance profile: plug in, disable aggressive battery saver, but avoid max turbo if hot; monitor temps.\n3. Reduce workload: smaller models/quantized/batched inference, lower resolution/context, close background apps.\n4. Ergonomics: external keyboard/mouse, screen height, breaks,"
        },
        {
          "prompt": "code",
          "run": 1,
          "elapsed_seconds": 9.964826999999786,
          "prompt_tokens": 50,
          "generation_tokens": 256,
          "prompt_tps": 134.21584726767261,
          "generation_tps": 26.713656380594895,
          "peak_memory_gb_reported": 21.55935528,
          "peak_memory_bytes_mx": 21559355280,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for a complete Python CLI program converting decimal GB and binary GiB. This is benign. Need provide code. Need ensure correctness: 1 GB decimal = 1e9 bytes; 1 GiB = 2^30 bytes. Conversions: GB -> GiB: bytes = GB * 1_000_000_000; GiB = bytes / 2**30. GiB -> GB: bytes = GiB * 2**30; GB = bytes / 1_000_000_000. Include argparse, validation positive finite numbers, two directions, error messages. Need maybe mention units. Provide complete program. No need for extra tools. Final should include code block and brief usage. Ensure no unsafe content. Need be concise but complete.\n\nDesign:\n- argparse with subcommands? Simpler: --direction {gb-to-gib,gib-to-gb}, --value float. Or positional value and --to. Use argparse.\n- Validate value: finite, >0? Could allow 0? Conversion of 0 valid. Say non-negative finite. Error messages via parser.error or sys.exit with message. Use custom exceptions.\n- Use Decimal"
        },
        {
          "prompt": "code",
          "run": 2,
          "elapsed_seconds": 9.817752874998405,
          "prompt_tokens": 50,
          "generation_tokens": 256,
          "prompt_tps": 134.87242224674452,
          "generation_tps": 27.124432913125485,
          "peak_memory_gb_reported": 21.559125904,
          "peak_memory_bytes_mx": 21559125904,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for a complete Python CLI program converting decimal GB and binary GiB. This is benign. Need provide code. Need ensure correctness: 1 GB decimal = 1e9 bytes; 1 GiB = 2^30 bytes. Conversions: GB -> GiB: bytes = GB * 1_000_000_000; GiB = bytes / 2**30. GiB -> GB: bytes = GiB * 2**30; GB = bytes / 1_000_000_000. Include argparse, validation positive finite numbers, two directions, error messages. Need maybe mention units. Provide complete program. No need for extra tools. Final should include code block and brief usage. Ensure no unsafe content. Need be concise but complete.\n\nDesign:\n- argparse with subcommands? Simpler: --direction {gb-to-gib,gib-to-gb}, --value float. Or positional value and --to. Use argparse.\n- Validate value: finite, >0? Could allow 0? Conversion of 0 valid. Say non-negative finite. Error messages via parser.error or sys.exit with message. Use custom exceptions.\n- Use Decimal"
        },
        {
          "prompt": "code",
          "run": 3,
          "elapsed_seconds": 9.926668709002115,
          "prompt_tokens": 50,
          "generation_tokens": 256,
          "prompt_tps": 135.13426771963802,
          "generation_tps": 26.812946186052788,
          "peak_memory_gb_reported": 21.559175056,
          "peak_memory_bytes_mx": 21559175056,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for a complete Python CLI program converting decimal GB and binary GiB. This is benign. Need provide code. Need ensure correctness: 1 GB decimal = 1e9 bytes; 1 GiB = 2^30 bytes. Conversions: GB -> GiB: bytes = GB * 1_000_000_000; GiB = bytes / 2**30. GiB -> GB: bytes = GiB * 2**30; GB = bytes / 1_000_000_000. Include argparse, validation positive finite numbers, two directions, error messages. Need maybe mention units. Provide complete program. No need for extra tools. Final should include code block and brief usage. Ensure no unsafe content. Need be concise but complete.\n\nDesign:\n- argparse with subcommands? Simpler: --direction {gb-to-gib,gib-to-gb}, --value float. Or positional value and --to. Use argparse.\n- Validate value: finite, >0? Could allow 0? Conversion of 0 valid. Say non-negative finite. Error messages via parser.error or sys.exit with message. Use custom exceptions.\n- Use Decimal"
        },
        {
          "prompt": "reasoning",
          "run": 1,
          "elapsed_seconds": 10.790776457997708,
          "prompt_tokens": 62,
          "generation_tokens": 256,
          "prompt_tps": 142.92057621300887,
          "generation_tps": 24.73918480794523,
          "peak_memory_gb_reported": 21.596781848,
          "peak_memory_bytes_mx": 21596781848,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user wants calculation, second method, common mistake. Need compute accurately.\n\nJourney: segment1 120 km in 1.5 h. segment2 80 km in 45 min = 0.75 h. Total distance = 200 km. Total time = 1.5 + 0.75 = 2.25 h. Average speed = 200 / 2.25 = 88.888... km/h = 88.9 km/h (or 800/9? 200/(9/4)=200*4/9=800/9=88.8889).\n\nSecond method: weighted average of segment speeds by time: speed1=120/1.5=80 km/h; speed2=80/0.75=106.666... km/h = 320/3? 80/(3/4)=80*4/3=320/3=106.6667. Weighted by time: (80*1.5 + (320/3)*0.75)/2.25. Numerator: 120 + (320/3)*(3/4)=120+80=200"
        },
        {
          "prompt": "reasoning",
          "run": 2,
          "elapsed_seconds": 9.919228375001694,
          "prompt_tokens": 62,
          "generation_tokens": 256,
          "prompt_tps": 143.95986748089672,
          "generation_tps": 27.005754743415576,
          "peak_memory_gb_reported": 21.595438156,
          "peak_memory_bytes_mx": 21595438156,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user wants calculation, second method, common mistake. Need compute accurately.\n\nJourney: segment1 120 km in 1.5 h. segment2 80 km in 45 min = 0.75 h. Total distance = 200 km. Total time = 1.5 + 0.75 = 2.25 h. Average speed = 200 / 2.25 = 88.888... km/h = 88.9 km/h (or 800/9? 200/(9/4)=200*4/9=800/9=88.8889).\n\nSecond method: weighted average of segment speeds by time: speed1=120/1.5=80 km/h; speed2=80/0.75=106.666... km/h = 320/3? 80/(3/4)=80*4/3=320/3=106.6667. Weighted by time: (80*1.5 + (320/3)*0.75)/2.25. Numerator: 120 + (320/3)*(3/4)=120+80=200"
        },
        {
          "prompt": "reasoning",
          "run": 3,
          "elapsed_seconds": 9.894284624999273,
          "prompt_tokens": 62,
          "generation_tokens": 256,
          "prompt_tps": 143.69982012364343,
          "generation_tps": 27.079132034207856,
          "peak_memory_gb_reported": 21.595668248,
          "peak_memory_bytes_mx": 21595668248,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user wants calculation, second method, common mistake. Need compute accurately.\n\nJourney: segment1 120 km in 1.5 h. segment2 80 km in 45 min = 0.75 h. Total distance = 200 km. Total time = 1.5 + 0.75 = 2.25 h. Average speed = 200 / 2.25 = 88.888... km/h = 88.9 km/h (or 800/9? 200/(9/4)=200*4/9=800/9=88.8889).\n\nSecond method: weighted average of segment speeds by time: speed1=120/1.5=80 km/h; speed2=80/0.75=106.666... km/h = 320/3? 80/(3/4)=80*4/3=320/3=106.6667. Weighted by time: (80*1.5 + (320/3)*0.75)/2.25. Numerator: 120 + (320/3)*(3/4)=120+80=200"
        }
      ],
      "weights_bytes": 19526043144,
      "draft_bytes_not_loaded": 635801416,
      "load_seconds": 4.926189665999118,
      "memory_after_load_bytes": 19526688776,
      "finished_at": "2026-11-21T16:41:51.837773+00:00",
      "medians": {
        "chat": {
          "prompt_tokens": 38,
          "generation_tokens": 256,
          "prompt_tps": 123.33694612191381,
          "generation_tps": 26.919339132396125,
          "peak_memory_gb_reported": 21.5198535,
          "peak_memory_bytes_mx": 21519853500
        },
        "code": {
          "prompt_tokens": 50,
          "generation_tokens": 256,
          "prompt_tps": 134.87242224674452,
          "generation_tps": 26.812946186052788,
          "peak_memory_gb_reported": 21.559175056,
          "peak_memory_bytes_mx": 21559175056
        },
        "reasoning": {
          "prompt_tokens": 62,
          "generation_tokens": 256,
          "prompt_tps": 143.69982012364343,
          "generation_tps": 27.005754743415576,
          "peak_memory_gb_reported": 21.595668248,
          "peak_memory_bytes_mx": 21595668248
        }
      }
    },
    {
      "schema": "modelfit.local-bench.v1",
      "model": "Inkling 2 Mini",
      "model_repo": "netprojects/Inkling-2-Mini-MLX-4-8bit",
      "revision": "67a7c329731b55ef3dc116ac697753201645bcd3",
      "model_path": "/Volumes/MINI2TO/models/huggingface/hub/models--netprojects--Inkling-2-Mini-MLX-4-8bit/snapshots/67a7c329731b55ef3dc116ac697753201645bcd3",
      "machine": {
        "model": "Mac17,16",
        "chip": "Apple M5 Pro",
        "ram_gb": 48,
        "arch": "arm64"
      },
      "software": {
        "python": "3.13.18",
        "mlx": "0.32.4",
        "mlx_vlm": "0.8.0",
        "mlx_vlm_source": "https://github.com/Neywa-Labs/mlx-vlm",
        "mlx_vlm_commit": "434a1bedd1e8a9ec9588462e112e6a0e58696872"
      },
      "settings": {
        "reasoning_effort": "none",
        "temperature": 0.0,
        "sampling": "greedy",
        "max_tokens": 256,
        "runs_per_prompt": 3,
        "prefill_step_size": 256,
        "draft_model": null,
        "offline": true
      },
      "command": "/Volumes/MINI2TO/scratch/dev-inkling2/venv/bin/python /Volumes/MINI2TO/scratch/dev-inkling2/bench.py --model /Volumes/MINI2TO/models/huggingface/hub/models--netprojects--Inkling-2-Mini-MLX-4-8bit/snapshots/67a7c329731b55ef3dc116ac697753201645bcd3 --output /Volumes/MINI2TO/scratch/dev-inkling2/bench-prefill256.json --runs 3 --max-tokens 256 --prefill-step-size 256",
      "prompts": {
        "chat": "Give five practical tips for keeping a laptop comfortable during long local AI inference runs. Explain each tip in one or two sentences.",
        "code": "Write a complete Python command-line program that converts between decimal gigabytes and binary gibibytes. Include argument parsing, input validation, two conversion directions, and useful error messages.",
        "reasoning": "A train travels 120 km in 1.5 hours, then 80 km in 45 minutes. Calculate its average speed for the whole journey. Show the calculation, check it with a second method, and explain one common mistake."
      },
      "started_at": "2026-11-21T16:42:59.013801+00:00",
      "status": "complete",
      "runs": [
        {
          "prompt": "chat",
          "run": 1,
          "elapsed_seconds": 10.526208833001874,
          "prompt_tokens": 38,
          "generation_tokens": 256,
          "prompt_tps": 78.29978204849265,
          "generation_tps": 27.134349078369283,
          "peak_memory_gb_reported": 21.511250533,
          "peak_memory_bytes_mx": 21511250533,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for five practical tips for keeping a laptop comfortable during long local AI inference runs, each explained in one or two sentences. Need answer concise. Need ensure safety? This is benign. Need provide practical tips: cooling/ventilation, power/thermal management, workload/batching, posture/ergonomics, monitoring/thermal throttling, maybe use external cooling/stand, reduce precision/quantization, close background apps, use battery/AC. Need exactly five tips, each 1-2 sentences. Mention comfort includes physical and thermal. Let's craft with five bullets. Ensure no overclaim. Mention local AI inference long runs: laptop gets hot, fans loud, battery drains, wrists/neck strain. Tips:\n1. Elevate and ventilate: laptop stand, clear vents, hard flat surface, avoid soft surfaces.\n2. Use AC and performance profile: plug in, disable aggressive battery saver, but avoid max turbo if hot; monitor temps.\n3. Reduce workload: smaller models/quantized/batched inference, lower resolution/context, close background apps.\n4. Ergonomics: external keyboard/mouse, screen height, breaks,"
        },
        {
          "prompt": "chat",
          "run": 2,
          "elapsed_seconds": 9.777056540999183,
          "prompt_tokens": 38,
          "generation_tokens": 256,
          "prompt_tps": 123.2856433765711,
          "generation_tps": 27.062319657500673,
          "peak_memory_gb_reported": 21.519493052,
          "peak_memory_bytes_mx": 21519493052,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for five practical tips for keeping a laptop comfortable during long local AI inference runs, each explained in one or two sentences. Need answer concise. Need ensure safety? This is benign. Need provide practical tips: cooling/ventilation, power/thermal management, workload/batching, posture/ergonomics, monitoring/thermal throttling, maybe use external cooling/stand, reduce precision/quantization, close background apps, use battery/AC. Need exactly five tips, each 1-2 sentences. Mention comfort includes physical and thermal. Let's craft with five bullets. Ensure no overclaim. Mention local AI inference long runs: laptop gets hot, fans loud, battery drains, wrists/neck strain. Tips:\n1. Elevate and ventilate: laptop stand, clear vents, hard flat surface, avoid soft surfaces.\n2. Use AC and performance profile: plug in, disable aggressive battery saver, but avoid max turbo if hot; monitor temps.\n3. Reduce workload: smaller models/quantized/batched inference, lower resolution/context, close background apps.\n4. Ergonomics: external keyboard/mouse, screen height, breaks,"
        },
        {
          "prompt": "chat",
          "run": 3,
          "elapsed_seconds": 9.760499707997951,
          "prompt_tokens": 38,
          "generation_tokens": 256,
          "prompt_tps": 123.22069718592323,
          "generation_tps": 27.109846408586808,
          "peak_memory_gb_reported": 21.520279484,
          "peak_memory_bytes_mx": 21520279484,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for five practical tips for keeping a laptop comfortable during long local AI inference runs, each explained in one or two sentences. Need answer concise. Need ensure safety? This is benign. Need provide practical tips: cooling/ventilation, power/thermal management, workload/batching, posture/ergonomics, monitoring/thermal throttling, maybe use external cooling/stand, reduce precision/quantization, close background apps, use battery/AC. Need exactly five tips, each 1-2 sentences. Mention comfort includes physical and thermal. Let's craft with five bullets. Ensure no overclaim. Mention local AI inference long runs: laptop gets hot, fans loud, battery drains, wrists/neck strain. Tips:\n1. Elevate and ventilate: laptop stand, clear vents, hard flat surface, avoid soft surfaces.\n2. Use AC and performance profile: plug in, disable aggressive battery saver, but avoid max turbo if hot; monitor temps.\n3. Reduce workload: smaller models/quantized/batched inference, lower resolution/context, close background apps.\n4. Ergonomics: external keyboard/mouse, screen height, breaks,"
        },
        {
          "prompt": "code",
          "run": 1,
          "elapsed_seconds": 9.913775708002504,
          "prompt_tokens": 50,
          "generation_tokens": 256,
          "prompt_tps": 133.41898832371191,
          "generation_tps": 26.8636558486182,
          "peak_memory_gb_reported": 21.559256848,
          "peak_memory_bytes_mx": 21559256848,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for a complete Python CLI program converting decimal GB and binary GiB. This is benign. Need provide code. Need ensure correctness: 1 GB decimal = 1e9 bytes; 1 GiB = 2^30 bytes. Conversions: GB -> GiB: bytes = GB * 1_000_000_000; GiB = bytes / 2**30. GiB -> GB: bytes = GiB * 2**30; GB = bytes / 1_000_000_000. Include argparse, validation positive finite numbers, two directions, error messages. Need maybe mention units. Provide complete program. No need for extra tools. Final should include code block and brief usage. Ensure no unsafe content. Need be concise but complete.\n\nDesign:\n- argparse with subcommands? Simpler: --direction {gb-to-gib,gib-to-gb}, --value float. Or positional value and --to. Use argparse.\n- Validate value: finite, >0? Could allow 0? Conversion of 0 valid. Say non-negative finite. Error messages via parser.error or sys.exit with message. Use custom exceptions.\n- Use Decimal"
        },
        {
          "prompt": "code",
          "run": 2,
          "elapsed_seconds": 9.959493291997205,
          "prompt_tokens": 50,
          "generation_tokens": 256,
          "prompt_tps": 133.59077772368474,
          "generation_tps": 26.733674552316124,
          "peak_memory_gb_reported": 21.560059792,
          "peak_memory_bytes_mx": 21560059792,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for a complete Python CLI program converting decimal GB and binary GiB. This is benign. Need provide code. Need ensure correctness: 1 GB decimal = 1e9 bytes; 1 GiB = 2^30 bytes. Conversions: GB -> GiB: bytes = GB * 1_000_000_000; GiB = bytes / 2**30. GiB -> GB: bytes = GiB * 2**30; GB = bytes / 1_000_000_000. Include argparse, validation positive finite numbers, two directions, error messages. Need maybe mention units. Provide complete program. No need for extra tools. Final should include code block and brief usage. Ensure no unsafe content. Need be concise but complete.\n\nDesign:\n- argparse with subcommands? Simpler: --direction {gb-to-gib,gib-to-gb}, --value float. Or positional value and --to. Use argparse.\n- Validate value: finite, >0? Could allow 0? Conversion of 0 valid. Say non-negative finite. Error messages via parser.error or sys.exit with message. Use custom exceptions.\n- Use Decimal"
        },
        {
          "prompt": "code",
          "run": 3,
          "elapsed_seconds": 9.83192254199821,
          "prompt_tokens": 50,
          "generation_tokens": 256,
          "prompt_tps": 134.86078128662146,
          "generation_tps": 27.086779920195017,
          "peak_memory_gb_reported": 21.559175056,
          "peak_memory_bytes_mx": 21559175056,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user asks for a complete Python CLI program converting decimal GB and binary GiB. This is benign. Need provide code. Need ensure correctness: 1 GB decimal = 1e9 bytes; 1 GiB = 2^30 bytes. Conversions: GB -> GiB: bytes = GB * 1_000_000_000; GiB = bytes / 2**30. GiB -> GB: bytes = GiB * 2**30; GB = bytes / 1_000_000_000. Include argparse, validation positive finite numbers, two directions, error messages. Need maybe mention units. Provide complete program. No need for extra tools. Final should include code block and brief usage. Ensure no unsafe content. Need be concise but complete.\n\nDesign:\n- argparse with subcommands? Simpler: --direction {gb-to-gib,gib-to-gb}, --value float. Or positional value and --to. Use argparse.\n- Validate value: finite, >0? Could allow 0? Conversion of 0 valid. Say non-negative finite. Error messages via parser.error or sys.exit with message. Use custom exceptions.\n- Use Decimal"
        },
        {
          "prompt": "reasoning",
          "run": 1,
          "elapsed_seconds": 9.897324541998387,
          "prompt_tokens": 62,
          "generation_tokens": 256,
          "prompt_tps": 143.91857016769967,
          "generation_tps": 27.06848063904365,
          "peak_memory_gb_reported": 21.595438156,
          "peak_memory_bytes_mx": 21595438156,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user wants calculation, second method, common mistake. Need compute accurately.\n\nJourney: segment1 120 km in 1.5 h. segment2 80 km in 45 min = 0.75 h. Total distance = 200 km. Total time = 1.5 + 0.75 = 2.25 h. Average speed = 200 / 2.25 = 88.888... km/h = 88.9 km/h (or 800/9? 200/(9/4)=200*4/9=800/9=88.8889).\n\nSecond method: weighted average of segment speeds by time: speed1=120/1.5=80 km/h; speed2=80/0.75=106.666... km/h = 320/3? 80/(3/4)=80*4/3=320/3=106.6667. Weighted by time: (80*1.5 + (320/3)*0.75)/2.25. Numerator: 120 + (320/3)*(3/4)=120+80=200"
        },
        {
          "prompt": "reasoning",
          "run": 2,
          "elapsed_seconds": 9.906306500000937,
          "prompt_tokens": 62,
          "generation_tokens": 256,
          "prompt_tps": 143.1718983580226,
          "generation_tps": 27.049239653021477,
          "peak_memory_gb_reported": 21.595420652,
          "peak_memory_bytes_mx": 21595420652,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user wants calculation, second method, common mistake. Need compute accurately.\n\nJourney: segment1 120 km in 1.5 h. segment2 80 km in 45 min = 0.75 h. Total distance = 200 km. Total time = 1.5 + 0.75 = 2.25 h. Average speed = 200 / 2.25 = 88.888... km/h = 88.9 km/h (or 800/9? 200/(9/4)=200*4/9=800/9=88.8889).\n\nSecond method: weighted average of segment speeds by time: speed1=120/1.5=80 km/h; speed2=80/0.75=106.666... km/h = 320/3? 80/(3/4)=80*4/3=320/3=106.6667. Weighted by time: (80*1.5 + (320/3)*0.75)/2.25. Numerator: 120 + (320/3)*(3/4)=120+80=200"
        },
        {
          "prompt": "reasoning",
          "run": 3,
          "elapsed_seconds": 9.910415166999883,
          "prompt_tokens": 62,
          "generation_tokens": 256,
          "prompt_tps": 143.81342759297635,
          "generation_tps": 27.032198661638905,
          "peak_memory_gb_reported": 21.595421772,
          "peak_memory_bytes_mx": 21595421772,
          "finish_reason": "length",
          "output_text": "<|message_model|><|content_thinking|>The user wants calculation, second method, common mistake. Need compute accurately.\n\nJourney: segment1 120 km in 1.5 h. segment2 80 km in 45 min = 0.75 h. Total distance = 200 km. Total time = 1.5 + 0.75 = 2.25 h. Average speed = 200 / 2.25 = 88.888... km/h = 88.9 km/h (or 800/9? 200/(9/4)=200*4/9=800/9=88.8889).\n\nSecond method: weighted average of segment speeds by time: speed1=120/1.5=80 km/h; speed2=80/0.75=106.666... km/h = 320/3? 80/(3/4)=80*4/3=320/3=106.6667. Weighted by time: (80*1.5 + (320/3)*0.75)/2.25. Numerator: 120 + (320/3)*(3/4)=120+80=200"
        }
      ],
      "weights_bytes": 19526043144,
      "draft_bytes_not_loaded": 635801416,
      "load_seconds": 4.6035802500009595,
      "memory_after_load_bytes": 19526688776,
      "finished_at": "2026-11-21T16:44:35.194731+00:00",
      "medians": {
        "chat": {
          "prompt_tokens": 38,
          "generation_tokens": 256,
          "prompt_tps": 123.22069718592323,
          "generation_tps": 27.109846408586808,
          "peak_memory_gb_reported": 21.519493052,
          "peak_memory_bytes_mx": 21519493052
        },
        "code": {
          "prompt_tokens": 50,
          "generation_tokens": 256,
          "prompt_tps": 133.59077772368474,
          "generation_tps": 26.8636558486182,
          "peak_memory_gb_reported": 21.559256848,
          "peak_memory_bytes_mx": 21559256848
        },
        "reasoning": {
          "prompt_tokens": 62,
          "generation_tokens": 256,
          "prompt_tps": 143.81342759297635,
          "generation_tps": 27.049239653021477,
          "peak_memory_gb_reported": 21.595421772,
          "peak_memory_bytes_mx": 21595421772
        }
      }
    }
  ],
  "notes": [
    "All 18 generations reached 256 tokens with finish_reason=length.",
    "Default and prefill_step_size=256 configurations were run separately, with one model load each.",
    "Prompt token counts were 38 (chat), 50 (code), and 62 (reasoning). These are short-prompt results, not 8K or 1M-context results.",
    "DSpark mtp.safetensors was present but not loaded.",
    "peak_memory_gb_reported is MLX-VLM/MLX-reported memory, not total macOS memory use.",
    "The first chat run of each configuration includes cold prefill and remains in the raw data; article tables use medians of three runs."
  ]
}
