Gemma is Google's family of open models, built from the same research and technology as the Gemini models, but with weights released so anyone can run it on their own machine. It comes in several sizes, from very small models that run on phones and laptops to larger ones for servers. Newer generations understand images (and audio in some sizes) as well as text, and support many languages. Its main difference from competitors is high quality for its size: on ordinary hardware you get results that used to be expected only from large models.

What is Gemma good for?#

  • Offline assistant: On a laptop with no internet, you summarize text, draft emails or edit writing.
  • Processing sensitive data: A clinic or law office classifies and summarizes confidential documents on its own machine.
  • AI features in a mobile app: An Android developer puts a small Gemma model inside an app so it works without a server.
  • Image description: It writes a text description from a product photo or screenshot.
  • Fine-tuning for a specific job: A technical team trains the model further on its own data to answer support questions.

Strengths and weaknesses#

Strengths:

  • High quality for its size, suited to ordinary hardware.
  • Better multilingual ability than most open models of the same size.
  • Permissive license in newer generations and support in all local-running tools.
  • Versions made for phones and low-power devices.

Weaknesses:

  • Doesn't match large cloud models at complex reasoning and heavy programming.
  • The very small sizes still make grammar and vocabulary mistakes in languages other than English.
  • Like any local model, its knowledge isn't up to date and it has no web access unless you connect it yourself.

Free version and pricing#

Gemma is free, and running it locally costs nothing beyond hardware. If you would rather not run it yourself, it is also available through Google's cloud services and other API providers at usage-based prices. Check the provider's own site for current prices.

Language support#

Among open models that run on a personal computer, Gemma is one of the better options for languages other than English: it understands them well and writes fluently in the medium and large sizes, though the very small sizes still make mistakes. You can get the model files through Ollama, LM Studio or Hugging Face, and after the download the model works completely offline.

Getting started in three steps#

  1. Install Ollama or LM Studio.
  2. Download a small or medium Gemma size that fits your system's RAM.
  3. Test it with a real piece of text from your own work, and if the quality isn't enough, download a larger size.

Gemma alternatives#

  • Llama: a bigger ecosystem, but weaker multilingual quality and no new generation.
  • Qwen: Alibaba's open models, strong at programming and reasoning tasks.
  • DeepSeek: strong open models whose distilled small versions run on a personal machine.
  • Ollama: the easiest way to run Gemma on your computer.

To choose a size and hardware, read running AI on a personal computer and open-source AI models: what they are and when they help.