Skip to content

Pick your style: Denkbaar

Book a conversation

Articles

History

Gemma 2: Google's open model fits on one GPU

·

Google released Gemma 2 in 9 and 27 billion parameters. The larger model runs on one data-centre GPU and scores like models twice its size.

A technician examines a single accelerator card in a server beside a screen with green bars.
The larger Gemma 2 model fits on one data-centre GPU. Illustration: Denkbaar.

What happened

On 27 June 2024 Google released Gemma 2 to researchers and developers. Gemma is Google’s family of open models, built with the same technology as Gemini. Gemma 2 came in two sizes: 9 billion and 27 billion parameters. Google wrote that the 27B model competes with models more than twice its size. It runs on one NVIDIA H100 GPU or one TPU host. That makes it cheaper to deploy.

Gemma 2 was available through Hugging Face, NVIDIA and Ollama.

Two model blocks of different sizes stand beside one accelerator card.
Gemma 2 came in two sizes. Illustration: Denkbaar.

Why it mattered then

According to Google, Gemma 2 reached a level that only closed models reached in December 2023. Half a year later you could run that yourself. This sped up the choice for open models in companies that do not want to send their data outside.

A worker with a headset tests a voice bot on a laptop beside a small server.
Choose a model based on the task, data and hardware. Illustration: Denkbaar.

What it means now

Gemma has grown since then. In November 2026 we use a Gemma model through Ollama in a trial with a Dutch voice bot. For a small company the question is not “open or closed”, but: which task, which data and which hardware? A model that fits on one GPU can run at a European host or on your own server. Always read the model’s licence terms before you use it for business.

Sources

  1. Google launches Gemma 2, its next generation of open models (Google)

Topics: open-weightsgooglegemma