Gemma
Small open-weight language models from Google, in 2B and 7B sizes
Gemma is a family of lightweight open language models released by Google DeepMind and other Google teams in March 2024. Built from the research and technology behind Gemini, it ships in 2 billion and 7 billion parameter sizes with pre-trained and instruction-tuned versions, and is meant to run on laptops, phones and cloud servers.
- Released 13 March 2024 in 2B and 7B parameter sizes
- Pre-trained and instruction-tuned checkpoints for each size
- Google reports wins over similar open models on 11 of 18 text tasks
- Built from research and technology behind Gemini
- Terms allow commercial use and distribution by any organisation
What it is
Gemma is a family of small open-weight language models from Google, released in March 2024. It comes in two sizes, 2 billion and 7 billion parameters, and each size has both a pre-trained and an instruction-tuned checkpoint. Google DeepMind and other teams across Google built it using research and technology from the Gemini programme. The company positions it for developers and researchers who want models they can run on a laptop, desktop, phone or cloud server, and fine-tune for tasks such as summarisation or retrieval-augmented generation.
How it came about
Google announced Gemma on 13 March 2024, with weights available worldwide from that day. The technical report, written by a Gemma Team of more than 100 authors led by Thomas Mesnard, appeared on arXiv the same day and was revised through April 2024. Google says the name comes from the Latin word for precious stone. The release was accompanied by a Responsible Generative AI Toolkit, toolchains for major frameworks, and free credits for researchers.
Inside the model
Google says Gemma shares technical and infrastructure components with Gemini, its largest general model. Both sizes ship as pre-trained and instruction-tuned variants. For the instruction-tuned versions, the company reports extensive fine-tuning plus reinforcement learning from human feedback, a method that trains the model against human preference judgements. Before training, automated techniques were used to strip certain personal and sensitive information from the datasets. Google states that benchmark data is excluded from training sets so that evaluations stay fair. The technical report covers performance, dataset composition and modelling methods in detail.
What it can do
Google reports that Gemma beats similarly sized open models on 11 of 18 text-based tasks, and says it surpasses significantly larger models on key benchmarks. The company claims best-in-class results for the models' sizes. The models are aimed at language understanding and reasoning tasks. Google notes that the models meet its standards for safe and responsible outputs, and points to its Model Card for the full evaluation results.
How to use it
Weights for both sizes are downloadable, with pre-trained and instruction-tuned checkpoints. Google's terms of use allow responsible commercial use and distribution by organisations of any size. Toolchains for inference and supervised fine-tuning run through Keras 3.0, PyTorch, JAX and Hugging Face Transformers, with integrations for MaxText, NVIDIA NeMo and TensorRT-LLM. Models run on laptops, desktops, IoT and mobile devices, and deploy on Vertex AI or Google Kubernetes Engine. Google offered free Kaggle and Colab access and cloud credits for new users and researchers.
Safety and testing
Google says it filtered personal and sensitive data from training sets using automated methods, and aligned the instruction-tuned models through fine-tuning and reinforcement learning from human feedback. It ran manual red-teaming, automated adversarial testing, and assessments of capabilities for dangerous activities, all described in the Model Card. A Responsible Generative AI Toolkit shipped alongside the weights, offering a method for building safety classifiers from few examples, a debugging tool for inspecting model behaviour, and best-practice guidance.
Reception and debate
Google frames the release as part of a longer record of open contributions, citing Transformers, TensorFlow, BERT, T5, JAX, AlphaFold and AlphaCode. The company argues that responsible release of large language models matters both for improving frontier model safety and for enabling further innovation. No independent reception is described here.
Why it matters
Google says it plans to expand the Gemma family with further variants for different applications. The models are positioned as a way for developers and researchers to build and fine-tune on their own data, with support across major frameworks and hardware platforms including NVIDIA GPUs and Google Cloud TPUs.
Questions people ask about Gemma
What is Gemma?
Gemma is a family of small open-weight language models from Google, released in March 2024. It comes in 2 billion and 7 billion parameter versions, each with a pre-trained and an instruction-tuned checkpoint, built from research and technology used for the Gemini models.
Who made it?
Google DeepMind together with other teams across Google. The technical report lists a Gemma Team of more than 100 authors, with Thomas Mesnard among the leads.
How do I use it?
Download the weights and run them on a laptop, desktop, phone or cloud server. Google provides toolchains for Keras 3.0, PyTorch, JAX and Hugging Face Transformers, plus deployment on Vertex AI and Google Kubernetes Engine.
Can I use it commercially?
Google's terms of use permit responsible commercial use and distribution for organisations of any size. The company also released a Responsible Generative AI Toolkit with safety classifiers, debugging tools and guidance.
How does it perform?
Google reports that Gemma outperforms similarly sized open models on 11 of 18 text-based tasks and surpasses significantly larger models on key benchmarks. Benchmark data is excluded from training sets, the company says.
Good reading and sources
There is no English Wikipedia article on Gemma yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2403.08295, blog.google/technology/developers/gemma-open-models/. 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.