Overview

Gemma is Google DeepMind's family of open models. Unlike Gemini, which you reach through Google's apps and API, Gemma's weights can be downloaded and run on your own hardware, from a cloud server down to a laptop or a phone. Google describes them as its most capable open models and pitches them at developers who want AI that runs wherever their users are.

The current line-up

The main release is Gemma 4, introduced in April 2026 and described by Google as purpose-built for advanced reasoning and agentic workflows. It comes in two groups:

E2B and E4B: small, efficient models aimed at mobile and IoT devices.
12B, 26B and 31B: larger models for advanced reasoning on personal computers. Google calls the 12B a unified, encoder-free multimodal model.

Around that core sits a growing list of specialised variants. Recent ones on the official page include EmbeddingGemma 2 (October 2026), which puts code, images, video and audio into one shared embedding space; DiffusionGemma, a text diffusion model built for generation speed; Gemma 4 QAT, compressed versions for phones and laptops; TranslateGemma, covering 55 languages; MedGemma for health AI development; FunctionGemma for function calling on devices; VaultGemma, a differentially private model; and T5Gemma encoder-decoder models. There's also Gemma Scope 2, a tool for safety researchers studying how the models behave.

How you use it, and what it costs

The quickest route is Google AI Studio, where you can try Gemma in the browser before downloading anything. From there, Google links to documentation, official variants and a community ecosystem it calls the "Gemmaverse" of fine-tuned versions and integrations. Google's Gemma pages list no price for the models themselves; what you pay for is the hardware or cloud compute you run them on. Read the Gemma licence terms on Google's site before shipping a commercial product. Checked October 2026 on deepmind.google/models/gemma and ai.google.dev/gemma.

Who it's for

Developers who need a model that works offline, on-device or inside their own infrastructure for privacy or cost reasons, and researchers who want to fine-tune a capable base model. If you just want a chat assistant, use Gemini instead.

Pros: open weights in sizes from phone to workstation; many task-specific variants; easy browser testing in AI Studio.
Cons: you handle deployment, updates and safety tuning yourself; the larger models need serious hardware; the number of variants can be confusing at first.