A beautiful, interactive CLI for benchmarking Ollama & LM Studio LLM inference speed

---

📸 Screenshots

  

✨ Features

• 🎯 Interactive Model Selection — Beautiful TUI multi-select with filtering by size and name
• 📊 Rich Metrics — Generation TPS, prompt eval TPS, TTFT, ITL, and model load time
• 🎨 Styled Terminal Output — Powered by Lipgloss and Huh
• 🔁 Multiple Epochs — Run N benchmarks per model, keep the best result
• 📝 Configurable Prompts — Code generation, chat, long-form, or custom file-based prompts
• 📈 Export Formats — CSV, JSON, or Go benchstat-compatible output
• 🔥 Batch Mode — Benchmark all matching models without interactive selection
• 🌡️ Automatic Warmup — Every model gets a lightweight warmup run + 1s settling delay before measurement (Ollama and LM Studio)
• 📈 Delta vs Last Run — Each result shows ▲/▼ change vs the previous stored benchmark for that model
• 📉 Stability Analysis — See variance, min/max across epochs
• 🧠 Thinking Mode — Optional --think flag sends "think": true to Ollama and LM Studio
• 🌐 Web Dashboard — Serve results over HTTP with lmspeedtest serve; chart bars show context size
• 🔐 Auth Support — Bearer token authentication for remote Ollama/LM Studio instances
• 🖥️ Remote-Friendly — Works with local or remote Ollama/LM Studio servers
• 🏢 Multi-Server Support — Manage and benchmark multiple Ollama/LM Studio servers
• 🔍 Update Checker — Check for new versions with a single command
• 🐚 Shell Completions — Bash, Zsh, and Fish completion scripts out of the box
• 📋 JSON Output — Machine-readable output for models, dashboard, info, and compare commands

---

🚀 Installation

Pre-built Binaries (Recommended)

Download the latest release for your platform from the releases page.

macOS (Intel)

curl -L https://github.com/notfixingit3/lmspeedtest/releases/latest/download/lmspeedtest_$(curl -s https://api.github.com/repos/notfixingit3/lmspeedtest/releases/latest | grep tag_name | cut -d '"' -f 4)_darwin_amd64.tar.gz | tar xz
mv lmspeedtest /usr/local/bin/

macOS (Apple Silicon)

curl -L https://github.com/notfixingit3/lmspeedtest/releases/latest/download/lmspeedtest_$(curl -s https://api.github.com/repos/notfixingit3/lmspeedtest/releases/latest | grep tag_name | cut -d '"' -f 4)_darwin_arm64.tar.gz | tar xz
mv lmspeedtest /usr/local/bin/

Linux (x86_64)

curl -L https://github.com/notfixingit3/lmspeedtest/releases/latest/download/lmspeedtest_$(curl -s https://api.github.com/repos/notfixingit3/lmspeedtest/releases/latest | grep tag_name | cut -d '"' -f 4)_linux_amd64.tar.gz | tar xz
sudo mv lmspeedtest /usr/local/bin/

Linux (ARM64)

curl -L https://github.com/notfixingit3/lmspeedtest/releases/latest/download/lmspeedtest_$(curl -s https://api.github.com/repos/notfixingit3/lmspeedtest/releases/latest | grep tag_name | cut -d '"' -f 4)_linux_arm64.tar.gz | tar xz
sudo mv lmspeedtest /usr/local/bin/

Windows (x86_64) Download lmspeedtest_*_windows_amd64.zip from the releases page, extract, and add to your PATH.

Windows (ARM64 - Snapdragon/Qualcomm) Download lmspeedtest_*_windows_arm64.zip from the releases page, extract, and add to your PATH.

Via Go Install

go install github.com/notfixingit3/lmspeedtest@latest

From Source

git clone https://github.com/notfixingit3/lmspeedtest.git
cd lmspeedtest
go build -o lmspeedtest
chmod +x lmspeedtest

---

🎬 Quick Start

# Configure your Ollama/LM Studio host
./lmspeedtest connect

# List available models
./lmspeedtest models

# Benchmark models ≤ 8 GB with interactive selection
./lmspeedtest test 8

# View your results dashboard
./lmspeedtest dashboard

---

📖 Usage

Commands

Command	Description
connect	Configure Ollama/LM Studio host and optional auth token
connect --add <name>	Add a new server profile
connect --list	List all configured server profiles
connect --default <name>	Set the default server profile
connect --use <name>	Switch to a different server profile
connect --remove <name>	Remove a server profile
info	Show server version, host, and auth status
doctor	Run diagnostics: check config, connectivity, and permissions
models [max_gb] [name_filter]	List models with metadata (params, quantization)
test <max_gb> [opts]	Benchmark matching models
dashboard [name_filter]	Show latest results per model, optionally filtered by name
compare <query>	Compare results — partial name match across all stored models
completions [shell]	Generate shell completion scripts (bash, zsh, fish)
export [--format fmt]	Export results (csv, json, benchstat, markdown)
prune	Remove results for models no longer on the active server
reset	Clear all benchmark results
serve [--port N]	Start web dashboard (default: 8080)
update	Check for available updates
--version	Show version information

Test Options

./lmspeedtest test 8                       # Interactive TUI selection
./lmspeedtest test 8 64k                   # Use 64k context window
./lmspeedtest test 8 32k qwen              # Filter by name + context
./lmspeedtest test 8 llama,gemma4          # Multi-filter: match any name
./lmspeedtest test 8 --all                 # Benchmark all (skip TUI)
./lmspeedtest test 8 --epochs 3            # Run 3 epochs, keep best
./lmspeedtest test 8 --template code       # Code generation prompt
./lmspeedtest test 8 --template chat       # Short chat prompt
./lmspeedtest test 8 --template long       # Long-form writing (default)
./lmspeedtest test 8 --prompt-file path.txt  # Custom prompt from file
./lmspeedtest test 8 --think               # Enable thinking/reasoning mode

Note: Thinking mode (--think) sends "think": true to Ollama and LM Studio. Warmup runs always use think: false regardless of this flag.

Doctor Exit Codes

The doctor command returns specific exit codes for programmatic use:

Code	Meaning
0	Pass — all checks OK
1	Warnings — potential issues detected
2	Config errors — invalid or missing configuration
3	Connectivity — cannot reach Ollama/LM Studio
4	Permissions — file or directory permission issues
5	Data errors — corrupted or unreadable results data

Real-World Examples

# Compare all your models at once
./lmspeedtest test 100 --all

# Deep benchmark: 5 epochs with long-form prompt
./lmspeedtest test 16 --epochs 5 --template long

# Quick code generation benchmark
./lmspeedtest test 8 --template code --epochs 3

# Deep benchmark with custom prompt
./lmspeedtest test 16 --epochs 5 --prompt-file prompt.txt

# Multi-filter: benchmark llama and gemma models only
./lmspeedtest test 16 llama,gemma --epochs 3

# Export for statistical analysis
./lmspeedtest export --format benchstat > results.bench
benchstat results.bench

# Start web dashboard
./lmspeedtest serve
./lmspeedtest serve --port 3000

# Clear all results
./lmspeedtest reset

# Connect to remote Ollama/LM Studio with auth
./lmspeedtest connect
# Host: https://ollama.example.com or http://10.1.6.30:1234
# Token: sk-abc123
./lmspeedtest info

# Check for updates
./lmspeedtest update

# Generate shell completions
./lmspeedtest completions bash > /usr/local/etc/bash_completion.d/lmspeedtest
./lmspeedtest completions zsh > /usr/local/share/zsh/site-functions/_lmspeedtest
./lmspeedtest completions fish > ~/.config/fish/completions/lmspeedtest.fish

# Get JSON output for scripting
./lmspeedtest models --json
./lmspeedtest dashboard --json
./lmspeedtest info --json
./lmspeedtest compare llama3.2:latest --json

# Check version
./lmspeedtest --version

# Multi-server: add and switch between servers
./lmspeedtest connect --add desktop --host http://192.168.1.10:11434
./lmspeedtest connect --add laptop --host http://192.168.1.11:11434 --token sk-abc
./lmspeedtest connect --list
# → default (http://localhost:11434) [active]
# → desktop (http://192.168.1.10:11434)
# → laptop (http://192.168.1.11:11434)
./lmspeedtest connect --use desktop
./lmspeedtest test 8 --all
./lmspeedtest connect --use laptop
./lmspeedtest test 8 --all
./lmspeedtest compare llama3.2:latest
# Shows results from all servers with server column

# Partial/fuzzy compare — matches any model containing "qwen"
./lmspeedtest compare qwen

# Dashboard filtered by name
./lmspeedtest dashboard qwen

# Benchmark with thinking/reasoning mode enabled
./lmspeedtest test 8 qwen --think

---

📊 Metrics Explained

Metric	Description
Tokens/sec	Generation speed — output tokens per second
Prompt TPS	Input processing speed — prompt eval tokens per second
TTFT	Time to first token — load + prompt eval duration
ITL	Inter-token latency — time between consecutive tokens
Load Time	Time to load model weights into GPU/CPU memory
Stability	Variance across epochs (stddev, min, max)

---

⚙️ Configuration

State is stored in ~/.lmspeedtest/:

• config.json — Server profiles (host URL, auth token, active profile)
• results.json — Last 3 benchmark results per model per server (capped at 3)

Multi-Server Setup

Manage multiple Ollama/LM Studio servers with named profiles:

# Add servers
./lmspeedtest connect --add desktop --host http://192.168.1.10:11434
./lmspeedtest connect --add laptop --host http://192.168.1.11:11434 --token sk-abc

# List profiles
./lmspeedtest connect --list

# Switch active server
./lmspeedtest connect --use desktop

# Remove a profile
./lmspeedtest connect --remove laptop

Results are stored per-server, so you can benchmark the same model on different hardware and compare.

---

📋 Requirements

• Go 1.25.8+
• Ollama or LM Studio instance (local or remote)

---

🤝 Contributing

Contributions are welcome! The codebase is organized into logical files:

• main.go — Entry point and command routing

• types.go — Structs and data models

• config.go — Configuration persistence and migration

• styles.go — Terminal styling helpers

• commands_connect.go — Server profile management

• commands_models.go — Model listing and benchmarking

• benchmark.go — Core benchmark logic

• commands_other.go — Dashboard, compare, export, serve, reset

• New subcommand → add case in main() switch

• New API call → add helper in the relevant command file

• New TUI form → use huh.NewForm(...) pattern

Please ensure your code passes golangci-lint and gosec.

---

☕ Support

If you find this tool useful, consider buying me a coffee:

---

📄 License

MIT License — see LICENSE for details.

---

Made with a fucked-up back and extremely questionable sleep habits by @notfixingit3