Gemma 2
Open lightweight language models from Google DeepMind, 2B to 27B parameters
Gemma 2 is a family of open language models released by Google DeepMind in 2024, in 2 billion, 9 billion and 27 billion parameter sizes. Built from research behind the Gemini models, it uses a redesigned Transformer and knowledge distillation to deliver strong performance at practical sizes, and is released under the Gemma licence for research and commercial use.
- Sizes of 2B, 9B and 27B parameters
- 2B and 9B trained with knowledge distillation
- 27B runs at full precision on one H100 or A100 80GB GPU
- Released under the commercially friendly Gemma licence
- Interleaved local-global attention and group-query attention
What it is
Gemma 2 is a set of open language models from Google DeepMind, released in 2024 as the second generation of the Gemma family. The models come in 2 billion, 9 billion and 27 billion parameter versions, with the 2B size announced as an upcoming addition. They are general-purpose text models aimed at researchers and developers, designed to run efficiently on ordinary hardware such as a single GPU or TPU host. The company says the 27B version offers competitive alternatives to models more than twice its size.
How it came about
Gemma 2 was developed by the Gemma Team at Google DeepMind, a group of around 200 authors led by Morgane Riviere and colleagues. The technical report was submitted on 31 July 2024 and later revised. It followed the first Gemma release and related variants such as CodeGemma, RecurrentGemma and PaliGemma. Google says the first Gemma launch passed 10 million downloads before Gemma 2 arrived.
Inside the model
The models use a modified Transformer design. The team interleaved local attention, which looks at nearby tokens, with global attention across the whole input, and used group-query attention to speed up inference. The 2B and 9B versions were trained with knowledge distillation, learning from a larger teacher model rather than predicting the next token directly. Google says the 27B model can run at full precision on one Google Cloud TPU host, an NVIDIA A100 80GB or an NVIDIA H100 GPU.
What it can do
Google reports that the 27B model gives the best performance in its size class and competes with models more than twice as large. The 9B version is said to outperform Llama 3 8B and other open models of similar size. The company frames this as performance previously only available from proprietary systems as of December. Detailed benchmark breakdowns are in the technical report.
How to use it
Weights are downloadable from Kaggle and Hugging Face Models, with Vertex AI Model Garden planned. The models run in Google AI Studio, and free access is offered through Kaggle and a Colab free tier. They work with Hugging Face Transformers, JAX, PyTorch, TensorFlow via Keras 3.0, vLLM, Gemma.cpp, Llama.cpp and Ollama, and are optimised for NVIDIA TensorRT-LLM. The Gemma licence allows sharing and commercial use.
Safety and testing
Google says it filtered pre-training data and ran testing against a set of metrics to find and reduce bias and risk, publishing results on public safety and representational-harm benchmarks. It provides a Responsible Generative AI Toolkit and the LLM Comparator for evaluation, and says it is working to open-source SynthID text watermarking for Gemma models.
Reception and debate
The first Gemma release drew more than 10 million downloads, according to Google, and inspired projects such as Navarasa, a model built around India's linguistic diversity. No independent reception details are given here.
Why it matters
Gemma 2 extends the Gemma line toward practical, low-cost deployment, running large-model-class output on a single accelerator. Google says it will keep developing specialised Gemma variants for more tasks, with a 2.6B Gemma 2 model announced as upcoming.
Questions people ask about Gemma 2
What is Gemma 2?
A family of open language models from Google DeepMind, released in 2024 in 2B, 9B and 27B parameter sizes. They are general-purpose text models meant for researchers and developers, built from research used in the Gemini models and released under the Gemma licence.
How is it different from the first Gemma?
Google says Gemma 2 uses a redesigned architecture with interleaved local and global attention plus group-query attention, and trains the smaller models by knowledge distillation. It is described as higher-performing and more efficient at inference than the first generation.
What hardware do I need?
The 27B model is designed to run at full precision on a single Google Cloud TPU host, an NVIDIA A100 80GB GPU or an NVIDIA H100 GPU. Smaller versions and quantised builds run on CPUs and consumer NVIDIA RTX hardware.
Can I use it commercially?
Yes. Google releases Gemma 2 under its commercially friendly Gemma licence, which the company says lets developers and researchers share and commercialise their work.
Where can I try or download it?
Weights are on Kaggle and Hugging Face Models, and the model can be tested in Google AI Studio. Free access is available through Kaggle and a Colab free tier, with Vertex AI deployment planned.
Good reading and sources
There is no English Wikipedia article on Gemma 2 yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2408.00118, blog.google/technology/developers/google-gemma-2/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.