llm-benchmark (ollama-benchmark)

LLM Benchmark for Throughput via Ollama (Local LLMs)

Measure how fast your local LLMs really are—with a simple, cross-platform CLI tool that tells you the tokens-per-second truth.

Installation prerequisites

Working Ollama installation.

To create a virtual environment via python3 -m venv

python3 -m venv .venv
## On Linux and macOS
source .venv/bin/activate
## On Windows Powershell or Cmd
.\.venv\Scripts\activate

To create a virtual environment via uv (For uv virtual environments (recommended for Python 3.13))

uv venv .venv --python 3.13
## On Linux and macOS
source .venv/bin/activate
## On Windows Powershell or Cmd
.\.venv\Scripts\activate

Installation Steps

Depending on your python setup either

pip install llm-benchmark

or

pipx install llm-benchmark

or uv

uv pip install llm-benchmark

Usage for general users directly

llm_benchmark run

Installation and Usage in Video format

It's tested on Python 3.10 and above.

ollama installation with the following models installed

7B model can be run on machines with 8GB of RAM

13B model can be run on machines with 16GB of RAM

Usage explaination

On Windows, Linux, and macOS, it will detect memory RAM size to first download required LLM models.

When memory RAM size is greater than or equal to 4GB, but less than 7GB, it will check if gemma:2b exist. The program implicitly pull the model.

ollama pull deepseek-r1:1.5b
ollama pull gemma:2b
ollama pull phi:2.7b
ollama pull phi3:3.8b

When memory RAM size is greater than 7GB, but less than 15GB, it will check if these models exist. The program implicitly pull these models

ollama pull phi3:3.8b
ollama pull gemma2:9b
ollama pull mistral:7b
ollama pull llama3.1:8b
ollama pull deepseek-r1:8b
ollama pull llava:7b

When memory RAM size is greater than 15GB, but less than 31GB, it will check if these models exist. The program implicitly pull these models

ollama pull gemma2:9b
ollama pull mistral:7b
ollama pull phi4:14b
ollama pull deepseek-r1:8b
ollama pull deepseek-r1:14b
ollama pull llava:7b
ollama pull llava:13b

When memory RAM size is greater than 31GB, it will check if these models exist. The program implicitly pull these models

ollama pull phi4:14b
ollama pull deepseek-r1:14b
ollama pull gpt-oss:20b

Python Poetry manually(advanced) installation

https://python-poetry.org/docs/#installing-manually

For developers to develop new features on Windows Powershell or on Ubuntu Linux or macOS

python3 -m venv .venv
. ./.venv/bin/activate
pip install -U pip setuptools
pip install poetry

Usage in Python virtual environment

poetry shell
poetry install
llm_benchmark hello jason

Example #1 send systeminfo and benchmark results to a remote server

llm_benchmark run

Example #2 Do not send systeminfo and benchmark results to a remote server

llm_benchmark run --no-sendinfo

Example #3 Benchmark run on explicitly given the path to the ollama executable (When you built your own developer version of ollama)

llm_benchmark run --ollamabin=~/code/ollama/ollama

Benchmark prompts are passed after -- so option-like text remains prompt data. In a shell wrapper, forward arguments with "$@" to preserve quoting and whitespace.

Example #4 run custom benchmark models

1. Create a custom benchmark file like following yaml format, replace with your own benchmark models, remember to use double quote for your model name

file_name: "custombenchmarkmodels.yml"
version: 2.0.custom
models:
  - model: "deepseek-r1:1.5b"
  - model: "qwen:0.5b"

2. run with the flag and point to the path of custombenchmarkmodels.yml

llm_benchmark run --custombenchmark=path/to/custombenchmarkmodels.yml

Custom benchmarks use models already installed in Ollama without pulling them from a registry. Models that are not installed are pulled as usual. The default benchmark keeps updating its configured models.

Reference

Ollama