Gemma 4 Models:
Complete Specs & Comparison
Google's latest open model family delivers best-in-class performance per parameter. Four powerful variants — from 2B embedded models to 31B dense — all under the permissive Apache 2.0 license.
What is Gemma 4?
Gemma 4 is Google DeepMind's newest open model family, released in early 2026 under the Apache 2.0 license. Named after the Latin word for "gem," this model family represents Google's most advanced open AI technology to date.
The standout achievement: Gemma 4 achieved #3 on Arenas AI ranking, placing it among the world's top AI models. What makes this remarkable is the efficiency — using significantly fewer parameters than competitors while delivering superior performance per parameter.
Available in four variants spanning from embedded devices (E2B at 2B parameters) to research-grade models (31B Dense), Gemma 4 brings multimodal capabilities — text, image, audio, and video — to the open model ecosystem for the first time. Whether you're building on a Raspberry Pi or a high-end GPU cluster, there's a Gemma 4 model optimized for your use case.
Model Specifications
Choose the right model for your hardware. From edge devices to enterprise servers, there's an optimal Gemma 4 variant.
| Model | Total Params | Active Params | Context Window | Modalities | Best For |
|---|---|---|---|---|---|
E2B | 5.1B | ~2.3B | 128K | Text / Image / Audio | Mobile, Edge devices |
E4B | 8.0B | ~4.5B | 128K | Text / Image / Audio | Mobile, Raspberry Pi |
26B MoE | 26B | ~3.8B | 256K | Text / Image / Video | Low-latency local use |
31B Dense | 31B | 31B | 256K | Text / Image / Video | Max quality, fine-tuning |
Built for everything
From embedded devices to research environments, Gemma 4 delivers state-of-the-art capabilities across the entire model family.
Advanced Reasoning
State-of-the-art chain-of-thought reasoning for complex problem-solving. Tackle multi-step math problems, logical puzzles, and intricate analytical tasks with unprecedented accuracy.
Agentic Workflows
Built-in function calling, structured JSON output, and robust system instruction following. Deploy reliable AI agents that can interact with external tools and APIs seamlessly.
Code Generation
Expert-level code generation across 140+ programming languages. From quick scripts to complex applications, Gemma 4 serves as your capable local code assistant.
Vision & Audio
Native multimodal understanding processes images, audio, and video alongside text. Analyze diagrams, transcribe speech, and understand visual context in real-time.
Long Context
Process up to 256K tokens in a single context window. Summarize entire codebases, analyze lengthy documents, or maintain coherent conversations over vast amounts of information.
140+ Languages
Trained on an exceptionally diverse multilingual corpus, Gemma 4 communicates fluently across a vast array of languages, from English and Mandarin to Swahili and Gaelic.
Truly Open, No Strings Attached
- Commercial use allowed
- No royalties or fees
- Modification permitted
- Distribution allowed
- No attribution required
Use It Your Way
Unlike models with restrictive licenses, Gemma 4's Apache 2.0 license lets you build commercial products, modify the model weights, and distribute it however you like — all without paying royalties or even mentioning Google.
This makes Gemma 4 ideal for startups, researchers, enterprises, and hobbyists alike. Build a paid SaaS product, create a local code assistant, or fine-tune it for your specific domain — the choice is yours.
Gemma 4 vs Gemma 3
A significant leap forward in capability, efficiency, and scope. See how Gemma 4 compares to its predecessor.
| Feature | Gemma 3 | Gemma 4 |
|---|---|---|
| Context Window | 32K - 128K | 128K - 256K |
| Parameters | Up to 12B | Up to 31B |
| Architecture | Dense | Dense + MoE |
| Modalities | Text / Image | Text / Image / Audio / Video |
| Languages | 40+ | 140+ |
| Arenas AI Ranking | #8 | #3 |
| Performance/Param | Good | Best-in-class |
Ready to get started?
Follow our quickstart guide to run Gemma 4 locally in minutes. No complex setup required.