Gemma

7:20 Gemma is Google's family of open-weight AI models, free to download and run on your own hardware. The current generation is Gemma 4, announced on Apr 2, 2026 in four sizes, from E2B for phones to a 31B model for a single GPU (a fifth, the 12B, followed on Jun 3, 2026), and released under the Apache 2.0 license, a change from the custom Gemma terms of earlier versions. Every Gemma 4 model reads text and images, the small ones and the 12B also hear audio, and context runs to 256K tokens. Google says Gemma has been downloaded more than 400 million times (Apr 2026).

1:36 Pick Gemma for a small, capable model on a phone, laptop or one graphics card, with a permissive license and image and audio input. For bigger open models that rival closed ones, look at DeepSeek, Qwen or Mistral Large 4; for OpenAI's open reasoning models, gpt-oss. For Google's best models, use Gemini. See Run AI Locally.

Quick facts

Gemma 4 models and memory

Source: Gemma 4 model overview, read Nov 6, 2026. The figures cover the weights only, not the memory a long conversation adds. The 26B A4B still needs room for all 26 billion parameters although only about 4 billion run per word. By those figures, the 4-bit 31B or 26B suits a 24 GB graphics card, the 12B a 12 GB card, and E2B and E4B a phone (Google lists 1.1 GB and 2.5 GB for them on mobile).

At launch Google said the 31B ranked third among open models on the Arena text leaderboard and the 26B sixth (Arena data of Apr 1, 2026, as Google reported it). On Apr 16, 2026 Google added multi-token prediction versions for faster answers (release notes).

Earlier versions and the Gemma family

Source: Gemma releases, read Nov 6, 2026 (Google's release list dates Gemma 4 to Mar 31; its blog announced it Apr 2).

Google also makes specialized Gemma models: MedGemma for medical text and images (1.5 on Jan 13, 2026), TranslateGemma for translation (Jan 15, 2026), EmbeddingGemma for search and retrieval (EmbeddingGemma 2, Oct 6, 2026), FunctionGemma (270M) for calling functions on devices, ShieldGemma for safety filtering, and VaultGemma, trained with differential privacy, all on the release list.

How to run Gemma

5:31 Download the weights from Hugging Face or Kaggle, or run ollama run gemma4 in Ollama (the default E4B tag is a 6.6 to 9.5 GB download, the 31B 19 to 20 GB); Google lists day-one support in llama.cpp, vLLM, MLX and LM Studio. On Android and iPhone, Google's AI Edge Gallery app runs E2B and E4B on the device. In the cloud, Gemma runs on Vertex AI, Cloud Run and GKE, and Google AI Studio lets you try the 31B and 26B in a browser.

Gemma vs the alternatives

• Gemini: Google's closed models, much stronger, but only through Google's apps and API.
• Qwen: Alibaba's open models span more sizes, including very large ones; a common pick for coding.
• Llama: Meta's open models; Llama 4 is larger but older (Apr 2025) and its license has more conditions.
• gpt-oss: OpenAI's text-only reasoning models, 20b and 120b, also Apache 2.0.
• Mistral: Ministral 3 (3B, 8B, 14B) is the closest European match in size and license.

Frequently asked questions

Is Gemma free? Yes. Google's Gemma models are free to download and use, and Gemma 4 is under the Apache 2.0 license, which allows commercial use.

What is the latest Gemma model? Gemma 4, announced Apr 2, 2026 in E2B, E4B, 26B A4B and 31B sizes, with Gemma 4 12B Unified added Jun 3, 2026; EmbeddingGemma 2 (Oct 6, 2026) is the newest specialized model.

What is the difference between Gemma and Gemini? Gemini is Google's closed flagship family, used in the Gemini app and API; Gemma is a smaller family built from the same research whose weights anyone can download and run.

Can Gemma run on a phone? Yes. Gemma 4 E2B needs about 1.1 GB of memory on a phone and E4B about 2.5 GB, and Google's AI Edge Gallery app runs them offline.

What is Gemma 3n? Gemma 3n, released Jun 26, 2025, is the phone-first version of Gemma 3, in E2B and E4B sizes; Gemma 4's E2B and E4B models succeed it.

Sources

• Gemma 4: Byte for byte, the most capable open models, Google, Apr 2, 2026
• Gemma releases, Google AI for Developers, read Nov 6, 2026
• Gemma 4 model overview, Google AI for Developers, read Nov 6, 2026
• Gemma, Google DeepMind, read Nov 6, 2026
• google/gemma-4-31b-it, Hugging Face, read Nov 6, 2026
• What’s new in Gemma 4, Google for Developers, Apr 2, 2026 (video)
• Google Gemma 4 Tutorial - Run AI Locally for Free, Teacher's Tech, Apr 5, 2026 (video)