Try LuxiEdge on your own machine.

1 · LuxiBook · public demonstration

LuxiBook: deterministic option-book pricing with receipts

LuxiBook takes a CSV position book, prices it with European Black-Scholes / Black-76, returns five Greeks per position, and attaches a signed receipt to the run. You bring time to expiry, rate, and volatility. No live market feed, no implied-volatility solver, no VaR.

Downloads

There is no macOS GPU binary. The CUDA build needs a live NVIDIA driver. Binaries are unsigned: on macOS, right-click the file and choose Open the first time.

Run it

# Mac CPU
chmod +x luxi-book-macos-arm64
./luxi-book-macos-arm64 price --book example_book.csv --out report.csv --receipt receipt.json

# Linux CPU
./luxi-book-linux-x86_64 price --book example_book.csv --out report.csv --receipt receipt.json

# Linux CUDA
./luxi-book-linux-x86_64-cuda price --book example_book.csv --out report.csv --receipt receipt.json --mode gpu

Verified public example book (2026-08-16)

Running the published example_book.csv through the published v0.2.1 binaries produced the same output hash on H200, H100, A100, RTX 5090 and RTX 4090, on both the LuxiBook CPU and CUDA pricing paths. Internal LuxiBook measurement. All twenty receipts from the original matrix remain published in the repository.

Every row is one physical machine with its own install key. The published folder preserves all 20 signed receipts from the original two-book matrix. Each machine's compute capability is recorded in its receipt. Apple Silicon has a published binary and was covered by an earlier check, but it is not part of this matrix; there is no macOS GPU binary.

1b · LuxiSize · included with LuxiBook

LuxiSize: free offline position-sizing CLI

LuxiSize is the sizing calculator in the free LuxiBook download. It is an offline command-line tool for liquidation price, position sizing, and maximum dollar loss. Every calculation can carry an Ed25519 signed receipt (lxr1_…).

This remains a legacy demonstration, not a retail crypto company and not the Risk State Engine. It ships as its own binary for now. For the option-book pricer, use LuxiBook.

Binaries are unsigned, so expect the usual macOS Gatekeeper or Windows SmartScreen friction on first launch. The luxirisk-v0.2 release files and the binaries in the repository's luxirisk/dist/ folder carry the same version label but are different builds with different digests: check a release download against that release's SHA256SUMS, and a repository copy against luxirisk/dist/SHA256SUMS.

2 · Luxi inference replay demo · closed binary

Replay a frozen reasoning job on your GPU or CPU

Luxi replay demo 1.0.1 runs the frozen v27 inference engine with DeepSeek-R1-Distill-Qwen-7B w8a8. It replays a 32-token sampled reasoning job on an NVIDIA GPU or a CPU, checks it against the signed November 9 H100 NVL record, signs the record, verifies records offline, compares records from two machines token by token and, on a GPU, reruns the 136-answer suite against the signed November 9 H100 NVL records. It is a reproducibility demo, not a chat app or a speed benchmark. The engine source is private.

Downloads

Run it

curl -LO https://github.com/RegularJoe-CEO/LuxiDemo/raw/88ebbe60311070f292cb81ba07a7132242ca694a/downloads/luxi-replay-demo/luxi-replay-demo-1.0.1-linux-x86_64.tar.gz
echo "c2114aa7bd6c47822bdf94de2f7d77f1aa773a32dcc3268072cfb979c2e5c8b3  luxi-replay-demo-1.0.1-linux-x86_64.tar.gz" | sha256sum -c
tar xzf luxi-replay-demo-1.0.1-linux-x86_64.tar.gz
cd luxi-replay-demo-1.0.1-linux-x86_64
./install.sh --gpu              # or: ./install.sh --cpu
./luxi fetch-model              # 8.7 GB once, every file checked by sha256
./luxi demo --device gpu        # or: --device cpu (32-token replay vs the signed Nov 9 H100 NVL record)
./luxi full-suite --device gpu  # GPU only: 136 answers vs the signed Nov 9 H100 NVL records

Tested November 10, 2026 (internal, single runs): RTX A6000 (release 1.0.1), RTX 5090 and A40 (release 1.0.0, same engine files) each matched 136/136 answers and 546,114 tokens against the signed November 9 H100 NVL records; the 32-token replay matched the signed November 9 H100 NVL record on all three GPUs and on AMD EPYC 7543, Intel Xeon Gold 6342 and one sandbox CPU. This is a new build identity: it does not inherit the November 9 nine-GPU result, and the CPU path covers the short replay only. A GPU request never falls back to the CPU. Proprietary closed binary, free for evaluation; model weights (8.7 GB, MIT) download from Hugging Face.

3 · Earlier inference evidence

Inference measurements

The November 9, 2026 nine-GPU and seven-CPU results are on Inference. Earlier Luxi Inference prefill and decode measurements are on Evidence and Research. Independent TestFort, internal matched, and internal absolute results are kept separate. Board joules are GPU-board energy, not wall-plug. Prefill positions are not decode tokens.