What it is

Gemma is a family of lightweight open models from Google, built from the research and technology Google used to create the Gemini models. Unlike Gemini, which you reach through Google’s apps and API, Gemma’s weights are published. Developers can download them to run and fine-tune on their own hardware, or build them into phones and IoT devices.

Google positions Gemma as a way to run AI applications wherever users need them, from cloud servers to laptops and phones. That is why the family focuses on intelligence per parameter and on compute and memory efficiency rather than sheer size.

Gemma has gone through several generations. According to Google DeepMind’s Gemma page, Gemma 4, released in April 2026, is “byte for byte, the most capable open models”, purpose-built for advanced reasoning and agentic workflows. It was followed by Gemma 4 12B, the quantisation-aware Gemma 4 QAT, and DiffusionGemma, a text diffusion model built on Gemma 4 and Gemini Diffusion research.

Key features

Several sizes. Gemma 4 includes E2B and E4B for phones and IoT devices, and 12B, 26B and 31B for personal computers and servers.

Multimodal and agentic. The gemma4 entry in the Ollama library lists vision, tools, thinking and audio. The small variants support a 128K context and the 12B and larger ones 256K. Google describes Gemma 4 12B as a unified, encoder-free multimodal model.

Many specialised variants. Around the core models Google has released task-specific variants, including:

  • MedGemma: medical text and image comprehension;
  • TranslateGemma: a collection of translation models covering 55 languages;
  • EmbeddingGemma: an embedding model designed for on-device use;
  • FunctionGemma: bespoke function calling on the edge;
  • ShieldGemma 2: a classifier for detecting policy-violating content;
  • T5Gemma: encoder-decoder Gemma models;
  • VaultGemma: a model trained with differential privacy;
  • Gemma Scope 2: interpretability tooling for studying model internals.

Broad ecosystem. Gemma is available on Kaggle and Hugging Face and works with Keras, PyTorch, JAX, Gemma.cpp, Ollama, LM Studio, Unsloth, Google AI Edge, Google Cloud and Android. Google calls the many community models derived from Gemma the “Gemmaverse”.

How it works

Shared roots with Gemini. Gemma shares its research and technical foundations with Gemini. The models are Transformer-based generative language models, scaled down so they can run on a single accelerator or even on-device.

Designed for efficiency. The family optimises at several levels:

  • Effective parameters in small variants. The “E” in E2B and E4B refers to effective parameter size. These variants target resource-limited devices such as phones and aim for as much capability as possible at low runtime cost.
  • Quantisation-aware training (QAT). Ordinary quantisation compresses weights to low precision after training and often costs some quality. QAT simulates low-precision arithmetic during training so the model adapts in advance and holds up better once compressed. Gemma 4 QAT is the compressed version aimed at phones and laptops.
  • Multi-token prediction drafters. Google released multi-token-prediction drafter models for Gemma 4, used for speculative decoding: a drafter quickly guesses the next several tokens and the main model verifies them in one pass, speeding up inference.

Diffusion-based text generation. DiffusionGemma generates text by diffusion rather than one token at a time, with the stated aim of maximising generation speed.

The flexibility of open weights. Because the weights are public, developers can do things that closed models do not allow: full or LoRA fine-tuning on their own data to learn domain vocabulary, quantising to smaller formats for memory-limited devices, deploying in fully isolated networks, or looking inside the model, as Gemma Scope 2 does, to study activations and behaviour. The many community models fine-tuned from Gemma are a direct result of that openness.

How to use it

Quickest way to try it: use Gemma in Google AI Studio with no local setup.

Run it locally with Ollama:

ollama run gemma4

Pick a size with a tag such as ollama run gemma4:e4b or ollama run gemma4:31b. You can then call the local API:

curl http://localhost:11434/api/chat -d '{
  "model": "gemma4",
  "messages": [{"role": "user", "content": "Introduce yourself in three sentences."}]
}'

Develop and fine-tune: download weights from Hugging Face or Kaggle and load them with Transformers, Keras or JAX. Tools such as Unsloth reduce memory needs for fine-tuning. For mobile deployment, see the Google AI Edge and Android documentation.

Read the terms of use on the model page before downloading and using the weights.

Use cases

  • On-device and offline apps: small variants on phones and laptops for features that work without a connection.
  • Private deployment: larger variants on your own servers for internal Q&A, document work and agent tasks.
  • Specialised domains: medical AI research with MedGemma, multilingual translation with TranslateGemma.
  • Retrieval: on-device embeddings with EmbeddingGemma for local RAG.
  • Content safety: screening generated content with ShieldGemma 2.
  • Safety and interpretability research: studying model internals with Gemma Scope 2.

FAQ

How is Gemma different from Gemini? Gemini is Google’s flagship model, used through the Gemini app and API, with weights not released. Gemma is a family of lightweight open models built from the same technology, with downloadable weights for local deployment and further development.

Which size suits my computer? Check the sizes listed in the Ollama library: E2B and E4B need a few GB, while 31B is close to 20GB. Start small if memory is limited.

Can I use Gemma commercially? Use of Gemma is governed by the terms published on its model pages, and variants may differ. Read the official terms before commercial use.

Can the models get things wrong? Yes. Google’s Gemma page warns that LLMs may produce inaccurate or offensive content and should not be relied on for medical, legal, financial or other professional advice.