Ollama Model Benchmark GUI

Start Here (20 Seconds)

1. Clone this repository
2. Run npm run setup:all to configure everything
3. Run npm run dev
4. Select your Ollama models, start the benchmark, and read results instantly

That is it. No heavy setup, no complex pipeline.

Fast benchmarking for Ollama models running locally (or on a remote endpoint), with a simple GUI, instant feedback, and practical metrics to quickly identify the fastest model on your real machine.

Why This Tool

This project has one clear goal: real tests, right away, without heavy setup.

• Lightweight and immediate Tkinter interface
• Multi-select Ollama models
• Speed ranking (tokens/sec)
• Optional hardware metrics (CPU, RAM, GPU when available)
• CSV export for comparisons and history
• Minimal dependencies, low footprint

Great for publishing benchmarks from real hardware without heavy pipelines.

Also ideal for Edge AI scenarios: small systems with limited resources (mini PCs, SBCs, embedded devices) where quickly understanding small-model performance in real conditions is critical.

Focus: Simplicity and Instant Use

• Quick start with npm commands
• Minimal configuration
• No external infrastructure required
• Designed for daily Linux desktop use
• Suitable for Edge AI devices with constrained resources

The tool is designed to add very little overhead on top of inference workload, so results better reflect real Ollama model performance.

This makes it useful for comparing quantizations and lightweight models when choosing the best balance between speed, RAM usage, and stability on limited hardware.

What It Measures

For each selected model:

• Avg tok/s (main metric)
• Std tok/s (run stability)
• Avg total s and Avg load s
• Avg CPU percent, Peak RAM MB
• Peak GPU percent, Avg GPU percent (best effort if nvidia-smi is available)

Quick start

1) Setup

npm run setup:all

This command:

1. checks Python
2. creates the local .venv
3. checks or installs GUI prerequisites (tkinter) when possible
4. validates project syntax

If you want to explicitly install system tkinter (with sudo), run:

npm run install:tk:system

2) Run

npm run dev

Alternative aliases:

npm run setup
npm run run

Use It in 30 Seconds

1. Enter your Ollama endpoint (default: http://localhost:11434)
2. Click Refresh models
3. Select the models
4. Click Start benchmark
5. Read ranking, row colors, and quick chart at a glance
6. Export results with Export CSV

Export CSV

Export generates two files:

• *_summary.csv with aggregated ranking
• *_runs.csv with run-by-run details

Perfect for public repositories, benchmark sharing, and follow-up analysis.

Important Notes

• tkinter is a system dependency (not a pip package), so Linux may require OS-level installation.
• With a remote Ollama endpoint, displayed hardware metrics come from the client machine running the GUI.

Stack

• Python 3
• Tkinter
• Ollama HTTP API (/api/tags, /api/generate)
• NPM scripts for a consistent workflow (setup and run)

Project Structure

• main.py: GUI entrypoint
• ui.py: interface, table, quick chart, CSV export
• ollama_client.py: Ollama HTTP client
• benchmark_runner.py: benchmark orchestration
• hw_monitor.py: CPU/RAM/GPU sampling
• models.py: result dataclasses

License

MIT. See LICENSE.