Skip to content
NexisNexis Wiki

Nexis Benchmark

Nexis Benchmark is a desktop app for benchmarking local AI models. Drop in ONNX or GGUF models, pick a task, run a standardized benchmark, and compare throughput, latency, memory, and accuracy side by side across inference backends.

Like the rest of the ecosystem, it’s built on Tauri 2 + React 19 + Vite + Tailwind v4: the Rust side owns the benchmark harness and backend abstraction, React owns model management and the charts.

Backends

Section titled “Backends”

The app has an end-to-end working pipeline. Three backends produce real measurements; the rest are deterministic simulations behind the same Rust Engine trait, so more real engines slot in without touching the harness, IPC, or UI.

Backend State
ONNX Runtime ✅ Real inference via the ort crate (prebuilt binaries, no cmake).
nexis-ml-rs ✅ Real wgpu/ndarray training throughput — spawns the nexis-ml-rs engine and reads its NDJSON stream.
llama.cpp ✅ Real GGUF inference via a prebuilt llama-bench (no cmake — it locates the binary).
Simulated ✅ Synthetic metrics with real event streaming, for UI and protocol testing.

Every result is labeled with its provenance in the UI: a green real badge vs sim, plus a per-run note — so simulated numbers can never be mistaken for measurements.

The benchmark protocol

Section titled “The benchmark protocol”

Each model × backend cell runs the same standardized sequence:

  1. Load the model.
  2. Warm-up — a configurable number of discarded runs.
  3. Measure — runs timed iterations.

The harness records per-run latency samples and derives tokens/sec, first-token latency, mean / p50 / p95 latency, and peak memory. Cancellation is cooperative — you can stop a running benchmark at any point with Esc.

Features

Section titled “Features”
  • Model library — drag-and-drop .onnx / .gguf files; format and task are detected from the file automatically.
  • Benchmark config — set the number of measured runs, warm-up runs, and token counts per protocol.
  • Results dashboard — a comparison chart, a live run matrix that fills in as results stream, and CSV export.
  • Browser mode — run the UI without Tauri (pnpm dev); it falls back to an in-process simulator with seeded demo models, which makes UI iteration fast.

Keyboard shortcuts

Section titled “Keyboard shortcuts”
Key Action
Ctrl+Enter (⌘+Enter on macOS) Run benchmark
Esc Stop a running benchmark
t Toggle light / dark theme

Where it fits in the ecosystem

Section titled “Where it fits in the ecosystem”

Nexis Benchmark closes the loop on the ecosystem’s local-AI story: the Nexis terminal runs local models (LM Studio, MLX, Ollama), the ML Suite trains them, and Benchmark tells you which model and backend is actually fastest on your hardware. It also exercises nexis-ml-rs directly as one of its real measurement engines.

Develop

Section titled “Develop”
Terminal window
pnpm install
pnpm tauri dev     # run the desktop app
pnpm dev           # browser-only mode with the simulator