{
  "as_of": "2026-10-19",
  "source": "Spark ~/handoffs/2026-10-19-site-inventory.md and .json (ledger L057). Every row carries a public link; a fact without a link is not on the page.",
  "judging": "Every number on this page is judged against the physical ceiling of the machine it ran on, and against my own previous run. A comparison with another project is allowed when it is honest, dated, and backed by a receipt or that project's own published numbers. Overclaims stay off this page. End to end means wall time through the server as a client sees it, prompt reading included. Chip time means the GPU's own time per generated token, decode only, with prompt reading excluded. For the 35B, the gap between chip time and end-to-end time comes from prompt reading. The two are labelled and never mixed.",
  "ceilings": [
    {
      "id": "qwen38-27b-fp8-gb10",
      "model": "Qwen3.8-27B FP8",
      "machine": "Spark GB10",
      "tok_s": 10.15,
      "how": "Memory ceiling, one token at a time: 273 GB/s nominal divided by 26,893,352,960 bytes of weights read per token. Arithmetic, not a measurement.",
      "link": "https://github.com/aien-dev/aien-sovereign-core-mv/blob/alpha/ceiling-and-model/docs/qwen38/CEILING.md",
      "receipt": "docs/qwen38/CEILING.md, calc:9822c2ba",
      "branch": "alpha/ceiling-and-model",
      "on_main": false
    }
  ],
  "machines": [
    {
      "id": "spark",
      "name": "NVIDIA DGX Spark",
      "detail": "Grace Blackwell GB10, 128 GB unified memory, AArch64",
      "role": "The development machine and the judge. Every speed receipt, every chip receipt, the verification pipeline and every agent run here.",
      "today": "Linux. Serves Qwen3.8-27B (FP8) through my own server, aien-serve-http, on my own C GPU kernels, built inside the Omega project. The AienOS C kernel has not booted on the Spark; a Rust-era kernel booted once on 2026-09-24.",
      "evidence_date": "2026-10-19",
      "evidence": "Live-server speed receipt, evidence/qwen38-serve-fast-20261019",
      "link": "https://github.com/aien-dev/aien-sovereign-core-mv/tree/qwen38/serve-fast/evidence/qwen38-serve-fast-20261019",
      "branch": "qwen38/serve-fast",
      "on_main": false,
      "more": [
        { "label": "First native boot, 2026-09-24", "link": "https://github.com/aien-dev/aienux/blob/main/evidence/m2_first_boot_2026-09-24.md" },
        { "label": "Fresh-account alpha test, 2026-10-19", "link": "https://github.com/aien-dev/aien-sovereign-core-mv/tree/alpha/fresh-account-evidence/evidence/alpha-fresh-account-20261019" }
      ]
    },
    {
      "id": "imac",
      "name": "Intel iMac 2017",
      "detail": "iMac18,3, Core i5-7500 with 4 cores, 16 GiB, AMD discrete GPU, Broadcom Ethernet",
      "role": "Bare metal AienOS, the x86_64 port of the C kernel. The whole screen is Aien's window: night sky, stars, black holes, Aien.",
      "today": "Boots the kernel from its own disk or netbooted from the Spark as a signed image, and sends a boot report back over the network. The kernel it runs today is in the open iMac pull request stack, not yet on main.",
      "evidence_date": "2026-10-17",
      "evidence": "iMac batch run: all cores, both disks read, Ethernet seen (evidence/imac_kernel1_2026-10-17, on main)",
      "link": "https://github.com/aien-dev/aienux/blob/main/evidence/imac_kernel1_2026-10-17",
      "branch": "main (evidence); kernel in PR #418 stack",
      "on_main": true,
      "more": [
        { "label": "The x86 port", "link": "https://github.com/aien-dev/aienux/blob/main/native/kernel/x86" },
        { "label": "PR #414 netboot", "link": "https://github.com/aien-dev/aienux/pull/414" },
        { "label": "PR #418 boot disk", "link": "https://github.com/aien-dev/aienux/pull/418" },
        { "label": "PR #424 screensaver", "link": "https://github.com/aien-dev/aienux/pull/424" }
      ]
    },
    {
      "id": "pi5",
      "name": "Raspberry Pi 5",
      "detail": "BCM2712, AArch64",
      "role": "Two jobs: the first real hardware for the AArch64 AienOS C kernel, and the project's house server on Linux.",
      "today": "Kernel runs are per test, from an SD image written for each run; no resident kernel is claimed. The house server runs readings-window, a read-only window onto the Spark's readings, and pi-register, the register of its services.",
      "evidence_date": "2026-10-13",
      "evidence": "Hardware runs: interrupts, four cores, user-mode tasks, scheduling, message passing, a log written to SD and read back (evidence/pi5_hardware_runs_2026-10-13, on main)",
      "link": "https://github.com/aien-dev/aienux/blob/main/evidence/pi5_hardware_runs_2026-10-13",
      "branch": "main",
      "on_main": true,
      "more": [
        { "label": "PR #345", "link": "https://github.com/aien-dev/aienux/pull/345" },
        { "label": "PR #346", "link": "https://github.com/aien-dev/aienux/pull/346" },
        { "label": "PR #402 ALLEN REFLEX on the Pi image (open)", "link": "https://github.com/aien-dev/aienux/pull/402" },
        { "label": "readings-window", "link": "https://github.com/aien-dev/readings-window" },
        { "label": "pi-register", "link": "https://github.com/aien-dev/pi-register" }
      ]
    },
    {
      "id": "qemu",
      "name": "QEMU",
      "detail": "AArch64 virt with AAVMF; x86_64 q35 with OVMF",
      "role": "The emulator path. Every AienOS kernel gate runs here first, and QEMU qualifies nothing physical, by the repository's own words.",
      "today": "Core-only and full AienOS kernel images for AArch64, the iMac kernel image for x86, the Atlas boot seed and the recovery images.",
      "evidence_date": "2026-10-09",
      "evidence": "24 of 24 emulator gates PASS at 7f0de7e (receipt on main). The gate script now lists 26 gates, so a newer receipt is owed.",
      "link": "https://github.com/aien-dev/aienux/blob/main/evidence/ck_gates_29ed03c97454354a18676f768273dd7f7a8e15462c4be1939d9994d277d0d639.json",
      "branch": "main",
      "on_main": true,
      "more": [
        { "label": "The gate list", "link": "https://github.com/aien-dev/aienux/blob/main/native/kernel/GATES.md" },
        { "label": "Atlas, the boot seed", "link": "https://github.com/aien-dev/atlas" }
      ]
    },
    {
      "id": "rig",
      "name": "Windows PC with an RTX 5070",
      "detail": "12 GB of GPU memory; the AIEN Rig target",
      "role": "The target for AIEN Rig: a local AI engine in a 12 GiB budget, a Windows frame-rate tuner, speech to text, a creator helper and one Windows app.",
      "today": "Nothing of mine has run on it yet. Every Rig artifact is built on GitHub Actions without a GPU, or measured on the Spark under a 12 GiB budget. There is no release tag yet.",
      "evidence_date": "2026-10-19",
      "evidence": "Rig engine receipts in a 12 GiB budget, measured on the Spark GB10, not on a 5070 (docs/inference/evidence/rig-engine-20261019)",
      "link": "https://github.com/aien-dev/aien-sovereign-core-mv/tree/rig/engine-12gb/docs/inference/evidence/rig-engine-20261019",
      "branch": "rig/engine-12gb",
      "on_main": false,
      "more": [
        { "label": "aien-rig", "link": "https://github.com/aien-dev/aien-rig" },
        { "label": "Omega for the 5070 (sm_120), on the Spark", "link": "https://github.com/aien-dev/omega-mv/tree/rig/gb20x/evidence/rig" }
      ]
    }
  ],
  "speeds": [
    {
      "id": "35b-e2e",
      "what": "End to end through a test server running a branch build, one conversation",
      "model": "Qwen3.6-35B-A3B FP8, a mixture-of-experts with about 3B parameters active per token",
      "machine": "Spark GB10",
      "engine": "C GPU stack (Omega project) behind aien-serve-http",
      "value": "22.0 to 22.5 tok/s",
      "detail": "21.95 and 22.46 tok/s in two runs: wall time from request out to last reply in, prompt reading included, greedy, at most 128 tokens per reply, on a test server running branch build bin-16b0e6f8 (port 18089, context 2048; the live server on port 18091 was stopped). Many conversations at once, in total: 57.1 tok/s at 8, 62.4 at 16, 65.7 to 66.4 at 32.",
      "tok_s": null,
      "ceiling": "",
      "previous": "",
      "date": "2026-10-21",
      "link": "https://github.com/aien-dev/aien-sovereign-core-mv/tree/9c62690a252ab1c0b75eb8933c550316b5c50d52/evidence/q36-batch-20261021",
      "receipt": "evidence/q36-batch-20261021, time-v2-16b0e6f8/time.jsonl",
      "branch": "main",
      "on_main": true,
      "verdict": "",
      "note": "The receipt commit is on main. Chip time for this model is decode only, prompt reading excluded, and is kept on the models page under its own label. The gap between that chip time and this end-to-end time comes from prompt reading."
    },
    {
      "id": "27b-live",
      "what": "Live server, chip time only (not end to end)",
      "model": "Qwen3.8-27B FP8",
      "machine": "Spark GB10",
      "engine": "C kernels (Omega project) behind aien-serve-http",
      "value": "7.63 tok/s",
      "detail": "131.05 ms of GPU time per token, median of 4 runs of 256 greedy tokens",
      "tok_s": 7.63,
      "ceiling": "qwen38-27b-fp8-gb10",
      "previous": "6.29 tok/s, the old binary, in the same receipt's before folder",
      "date": "2026-10-19",
      "link": "https://github.com/aien-dev/aien-sovereign-core-mv/tree/qwen38/serve-fast/evidence/qwen38-serve-fast-20261019",
      "receipt": "evidence/qwen38-serve-fast-20261019",
      "branch": "qwen38/serve-fast",
      "on_main": false,
      "verdict": "",
      "note": "Real traffic over 22 replies, chip median from the server log: 7.52 tok/s. The engine alone measures 123.0 ms; the 8 ms gap to the server is not explained yet."
    },
    {
      "id": "27b-fresh",
      "what": "Fresh account, public guide, new Linux user, chip time only",
      "model": "Qwen3.8-27B FP8",
      "machine": "Spark GB10, shared with other load",
      "engine": "C kernels (Omega project) behind aien-serve-http",
      "value": "7.657 tok/s",
      "detail": "130.6 ms of GPU time per token, median of 3 runs; answers match the shipped reference",
      "tok_s": 7.657,
      "ceiling": "qwen38-27b-fp8-gb10",
      "previous": "",
      "date": "2026-10-19",
      "link": "https://github.com/aien-dev/aien-sovereign-core-mv/tree/alpha/fresh-account-evidence/evidence/alpha-fresh-account-20261019",
      "receipt": "evidence/alpha-fresh-account-20261019/speed",
      "branch": "alpha/fresh-account-evidence",
      "on_main": false,
      "verdict": "install.sh failed as written; the server was built by hand before this run",
      "note": ""
    },
    {
      "id": "27b-engine",
      "what": "Engine alone, in process, chip time only",
      "model": "Qwen3.8-27B FP8",
      "machine": "Spark GB10",
      "engine": "C kernels (Omega project), engine test tool",
      "value": "8.13 tok/s",
      "detail": "123.04 ms per token, mean of 3 runs; 24 of 24 golden tokens",
      "tok_s": 8.13,
      "ceiling": "qwen38-27b-fp8-gb10",
      "previous": "6.40 tok/s, my run before this series",
      "date": "2026-10-19",
      "link": "https://github.com/aien-dev/omega-mv/tree/qwen38/decode-speed/evidence",
      "receipt": "omega evidence/q38-decode-speed-chip-20261019T*",
      "branch": "qwen38/decode-speed",
      "on_main": false,
      "verdict": "",
      "note": "Which of the four receipt folders holds the final three runs has not been marked yet."
    },
    {
      "id": "27b-fp8-head",
      "what": "FP8 final word-scoring step, A against B, chip time only",
      "model": "Qwen3.8-27B FP8",
      "machine": "Spark GB10",
      "engine": "C kernels (Omega project), engine test tool",
      "value": "8.452 tok/s",
      "detail": "118.31 ms per token with the FP8 head; 124.25 ms (8.048 tok/s) with the bf16 head",
      "tok_s": 8.452,
      "ceiling": "qwen38-27b-fp8-gb10",
      "previous": "8.048 tok/s with the bf16 head, same session",
      "date": "2026-10-19",
      "link": "https://github.com/aien-dev/aien-sovereign-core-mv/tree/qwen38/w4-decode/evidence/qwen38-ab-fp8head-w4-20261019T1230Z",
      "receipt": "evidence/qwen38-ab-fp8head-w4-20261019T1230Z",
      "branch": "qwen38/w4-decode",
      "on_main": false,
      "verdict": "FAILS the full-model gate on p95 drift, so it stays off",
      "note": ""
    },
    {
      "id": "27b-w4",
      "what": "4-bit weights, whole model, chip time only",
      "model": "Qwen3.8-27B MXFP4",
      "machine": "Spark GB10, chip shared with two servers",
      "engine": "C kernels (Omega project), engine test tool",
      "value": "72.6 to 76.2 ms per token",
      "detail": "about 13.1 to 13.8 tok/s; FP8 in the same session 119.35 ms. The receipt records milliseconds; the tok/s is arithmetic on it.",
      "tok_s": null,
      "ceiling": "",
      "previous": "",
      "date": "2026-10-19",
      "link": "https://github.com/aien-dev/omega-mv/tree/cascade/w4-gemv/evidence",
      "receipt": "omega evidence/cascade-w4-wm-*",
      "branch": "cascade/w4-gemv",
      "on_main": false,
      "verdict": "24 of 24 tokens identical, but FAILS the FP8 logit-drift gate, as 4-bit is expected to",
      "note": "Several rounds failed at load for lack of memory; the failures are published too."
    },
    {
      "id": "rig-8b-fp8",
      "what": "Rig engine in a 12 GiB budget, chip time only",
      "model": "Qwen3-8B FP8",
      "machine": "Spark GB10 under a 12 GiB device budget, not an RTX 5070",
      "engine": "Omega dense door",
      "value": "15.63 tok/s",
      "detail": "5 of 5 golden prompts identical",
      "tok_s": 15.63,
      "ceiling": "",
      "previous": "",
      "date": "2026-10-19",
      "link": "https://github.com/aien-dev/aien-sovereign-core-mv/tree/rig/engine-12gb/docs/inference/evidence/rig-engine-20261019/chip-fp8",
      "receipt": "rig-engine-20261019/chip-fp8",
      "branch": "rig/engine-12gb",
      "on_main": false,
      "verdict": "",
      "note": "Nothing has run on the 5070 yet."
    },
    {
      "id": "rig-8b-w4",
      "what": "Rig engine in a 12 GiB budget, 4-bit, chip time only",
      "model": "Qwen3-8B MXFP4",
      "machine": "Spark GB10 under a 12 GiB device budget, not an RTX 5070",
      "engine": "Omega dense door",
      "value": "20.17 tok/s",
      "detail": "5 of 5 golden prompts identical against the host reference; device memory 5.158 GiB; the swap-kv control fails as it must",
      "tok_s": 20.17,
      "ceiling": "",
      "previous": "15.63 tok/s with FP8 weights, same day",
      "date": "2026-10-19",
      "link": "https://github.com/aien-dev/aien-sovereign-core-mv/tree/rig/engine-12gb/docs/inference/evidence/rig-engine-20261019/chip-w4-ceil",
      "receipt": "rig-engine-20261019/chip-w4-ceil",
      "branch": "rig/engine-12gb",
      "on_main": false,
      "verdict": "",
      "note": ""
    },
    {
      "id": "27b-ttft",
      "what": "Cold time to first token at long context",
      "model": "Qwen3.8-27B FP8",
      "machine": "Spark GB10",
      "engine": "C kernels (Omega project) behind aien-serve-http",
      "value": "936.7 s at 128k tokens",
      "detail": "The 32k and 64k values are in the receipt",
      "tok_s": null,
      "ceiling": "",
      "previous": "",
      "date": "2026-10-18",
      "link": "https://github.com/aien-dev/aien-sovereign-core-mv/tree/rig/engine-12gb/evidence/qwen38-prefill-20261018T1836Z",
      "receipt": "evidence/qwen38-prefill-20261018T1836Z",
      "branch": "qwen38/prefill-ttft",
      "on_main": false,
      "verdict": "",
      "note": ""
    }
  ]
}