{
  "accessed": "2026-10-23",
  "scope": "Published results not reproduced; metric boundaries differ; no hardware speedup ratios",
  "rows": [
    {
      "table": 1,
      "columns": [
        "Release / workload",
        "Measured result",
        "Timing boundary / validation"
      ],
      "sources": []
    },
    {
      "table": 1,
      "columns": [
        "v124 · SmolVLM 384",
        "48.88 s / 51.36 s per caption",
        "Two saved images; 30/28 tokens; loading included. Original ONNX token agreement. Approximate features."
      ],
      "sources": []
    },
    {
      "table": 1,
      "columns": [
        "v125 · warm SmolVLM",
        "42.08 s / 41.00 s",
        "Same saved workloads as v124; setup excluded. Phase copies remain. Features and tokens bit-exact to v124."
      ],
      "sources": []
    },
    {
      "table": 1,
      "columns": [
        "v139 · SmolVLM2-256M 384",
        "33.83 s post-load; 59.39 s launch-to-caption",
        "Fresh 18-token caption including EOS; encoder 16.79 s. ONNX token agreement; strict FP32 feature tolerance fails."
      ],
      "sources": []
    },
    {
      "table": 1,
      "columns": [
        "v144 / v143 kernel · SmolLM-135M-Instruct",
        "7.11 s first token; 20.92 s for 32 tokens",
        "Loading and speech excluded. Saved FP32 reference32/32. About 2.25 output tok/s after first, derived as31/(20.917−7.112)."
      ],
      "sources": []
    },
    {
      "table": 1,
      "columns": [
        "v145 · LLM-jp-3-150m-instruct3",
        "2.21–2.53 s first token; ~5.8 output tok/s",
        "Paris/Osaka filtered controls,32 tokens, original FP32 reference. Speech and model loading excluded."
      ],
      "sources": []
    },
    {
      "table": 1,
      "columns": [
        "v145 · Japanese ASR",
        "Original cold:18.29 s → balanced:8.53 s; original warm repeats:1.93 /2.76 s",
        "Persistent recognizer. Cold profile comparison on identical synthesized PCM; warm values use original search. Not a human speech score. See the live-test panel below."
      ],
      "sources": []
    },
    {
      "table": 2,
      "columns": [
        "Device / backend",
        "Model / precision",
        "Published measurement",
        "Comparison limits / source"
      ],
      "sources": []
    },
    {
      "table": 2,
      "columns": [
        "Jetson Orin Nano Super · CUDA",
        "SmolLM-135M base · Q2_K",
        "190.0 benchmark tok/s",
        "Prompt + generated tokens divided by timed inference, not output-only decode. Four threads, five-run median, llama.cpp b11146,24 Sep2026. Base model differs from v144 Instruct. yzma raw report / metric code."
      ],
      "sources": [
        "https://github.com/hybridgroup/yzma/blob/main/benchmarks/linux.md",
        "https://github.com/hybridgroup/yzma/blob/main/pkg/llama/benchmark_test.go"
      ]
    },
    {
      "table": 2,
      "columns": [
        "Jetson Orin Nano Super · CPU",
        "SmolLM-135M base · Q2_K",
        "84.1 benchmark tok/s",
        "Same external suite as above; not measured with AiOnIc's precision or prompt."
      ],
      "sources": []
    },
    {
      "table": 2,
      "columns": [
        "Raspberry Pi4 Model B · CPU",
        "SmolLM-135M base · Q2_K",
        "35.4 benchmark tok/s",
        "Same external suite, four threads. Useful within that suite; no direct AiOnIc decode-rate ratio."
      ],
      "sources": []
    },
    {
      "table": 2,
      "columns": [
        "Arduino UNO Q · CPU",
        "SmolLM-135M base · Q2_K",
        "32.2 benchmark tok/s",
        "Same suite. CPU result, not an NPU measurement."
      ],
      "sources": []
    },
    {
      "table": 2,
      "columns": [
        "ROCK5B+ / RK3588 · three NPU cores",
        "SmolVLM2-256M · W8A8 decoder / FP16 vision,384²",
        "62.81 generated tok/s;97.59ms prefill",
        "78 prefill /166 generated tokens; RKLLM1.2.3; peak0.59GB reported. Vision/camera time is not included in the decoder number. Checkpoint revision and ONNX-equivalent accuracy not established. Radxa run log."
      ],
      "sources": [
        "https://docs.radxa.com/e/e54c/app-development/artificial-intelligence/rkllm-smolvlm2"
      ]
    },
    {
      "table": 2,
      "columns": [
        "Raspberry Pi5 · CPU",
        "Our exact SmolLM-135M-Instruct / LLM-jp150M configuration",
        "Not found in verified sources",
        "No same-checkpoint, same-precision, same-prompt measurement available. See the explicitly different-model context below."
      ],
      "sources": []
    },
    {
      "table": 2,
      "columns": [
        "Jetson Nano (legacy) / AGX Orin",
        "Our exact deployed configurations",
        "Not found in verified sources",
        "Do not substitute Orin Nano numbers for legacy Jetson Nano, or scale by TOPS."
      ],
      "sources": []
    },
    {
      "table": 3,
      "columns": [
        "Platform",
        "Different workload",
        "Published value",
        "Limits / source"
      ],
      "sources": []
    },
    {
      "table": 3,
      "columns": [
        "Raspberry Pi5 · Ollama CPU",
        "SmolLM1.7B, quantization not specified",
        "7.50 decode tok/s;11.73 prefill tok/s",
        "28 prompt /393 generated tokens. This is1.7B, not135M, and a different prompt. Firsthand2024 run: author's output."
      ],
      "sources": [
        "https://itsfoss.com/smollm-raspberry-pi/"
      ]
    },
    {
      "table": 3,
      "columns": [
        "Jetson Orin Nano8GB · Super",
        "SmolLM2-1.7B",
        "64.5 tok/s",
        "NVIDIA JetPack6.2 table4; larger different-generation model. NVIDIA vendor report."
      ],
      "sources": [
        "https://developer.nvidia.com/blog/nvidia-jetpack-6-2-brings-super-mode-to-nvidia-jetson-orin-nano-and-jetson-orin-nx-modules/"
      ]
    },
    {
      "table": 3,
      "columns": [
        "Jetson Orin NX16GB · Super",
        "SmolLM2-1.7B",
        "68.8 tok/s",
        "Same vendor source, table6; not the150M Japanese checkpoint."
      ],
      "sources": []
    },
    {
      "table": 3,
      "columns": [
        "Pi5 + DEEPX DX-M1M / DX-M1",
        "Qwen3-1.7B,96 prefill tokens",
        "4.64 /11.60 tok/s; TTFT599 /545ms",
        "Different model and toolchain; no matching SmolVLM/LLM-jp result. Raspberry Pi / Sixfab report."
      ],
      "sources": [
        "https://www.raspberrypi.com/news/seeing-understanding-and-responding-low-power-cnn-vlm-and-slm-workloads-on-raspberry-pi-5/"
      ]
    },
    {
      "table": 3,
      "columns": [
        "Hailo-10H / other edge NPUs",
        "Matching checkpoints and workload",
        "Not verified",
        "Supported-model lists or peak TOPS do not establish caption latency. A matching measured run is still needed."
      ],
      "sources": []
    },
    {
      "table": 4,
      "columns": [
        "Measurement",
        "v148 replay",
        "v149 live"
      ],
      "sources": []
    },
    {
      "table": 4,
      "columns": [
        "Whisper recognition",
        "20.991 s",
        "20.965 s"
      ],
      "sources": []
    },
    {
      "table": 4,
      "columns": [
        "Cached speech decoder",
        "3.867 s",
        "3.681 s"
      ],
      "sources": []
    },
    {
      "table": 4,
      "columns": [
        "Processing through playback completion",
        "45.143 s · sequential stage sum",
        "41.563 s · measured interval"
      ],
      "sources": []
    },
    {
      "table": 4,
      "columns": [
        "Question onset → spoken reply completed",
        "48.064 s · replay estimate",
        "44.483 s · estimated voice onset"
      ],
      "sources": []
    },
    {
      "table": 4,
      "columns": [
        "Whisper execution submissions",
        "451",
        "355"
      ],
      "sources": []
    },
    {
      "table": 5,
      "columns": [
        "Live question / gain",
        "ASR",
        "LLM",
        "Synthesis / playback",
        "Question end → audio start / complete",
        "Recognition"
      ],
      "sources": []
    },
    {
      "table": 5,
      "columns": [
        "Paris ·35%",
        "8.20s",
        "7.57s",
        "6.05 /12.67s",
        "22.84 /35.51s",
        "Failed: あるのはどこにありますか"
      ],
      "sources": []
    },
    {
      "table": 5,
      "columns": [
        "Osaka ·35%",
        "10.29s",
        "8.12s",
        "6.50 /13.67s",
        "25.96 /39.64s",
        "Failed: 教えて → 進んで"
      ],
      "sources": []
    },
    {
      "table": 5,
      "columns": [
        "Paris ·75%",
        "13.28s",
        "7.35s",
        "9.21 /19.19s",
        "30.91 /50.11s",
        "Failed: 今ありますか"
      ],
      "sources": []
    },
    {
      "table": 5,
      "columns": [
        "Osaka ·75%",
        "6.34s",
        "7.99s",
        "3.18 /6.55s",
        "18.56 /25.11s",
        "Matches requested wording"
      ],
      "sources": []
    }
  ]
}