Gemma 4
Apache 2.0 Open Source

Gemma 4 Open Models

Google DeepMind's most capable open-source AI model family. Multimodal understanding across text, image, video, and audio with up to 256K context and Apache 2.0 license.

Model Variants

Four purpose-built variants from edge devices to workstation-grade performance, all under the Apache 2.0 license.

Dense

Gemma 4 E2B

2B Parameters · 128K Context

Ultra-lightweight model optimized for on-device and edge deployments. Delivers strong performance in a compact footprint suitable for mobile and IoT applications.

Dense

Gemma 4 E4B

4B Parameters · 128K Context

Balanced model offering excellent quality-to-size ratio. Ideal for laptop and desktop deployments where resources are limited but high-quality output is required.

Mixture of Experts (128 Experts)

Gemma 4 26B A4B

26B Total / 4B Active Parameters · 256K Context

Sparse Mixture-of-Experts architecture with 128 experts, activating only 4B parameters per inference. Achieves large-model quality with small-model compute cost.

Dense

Gemma 4 31B

31B Parameters · 256K Context

Flagship dense model delivering state-of-the-art performance across all benchmarks. Best choice when maximum quality and reasoning depth are the priority.

Capabilities

Gemma 4 sets a new standard for open-source AI with multimodal understanding, extended context, and agentic abilities.

🎨

Multimodal Understanding

Natively processes text, images, video, and audio in a single model. Understand documents, analyze charts, describe scenes, and transcribe speech without separate pipelines.

📄

Long Context Window

Up to 256K token context enables processing entire codebases, lengthy documents, and extended conversations without losing coherence or critical details.

🤖

Agentic Capabilities

Built-in function calling and structured JSON output enable autonomous tool use, multi-step planning, and integration with external APIs and services.

💻

Code Generation

Achieves 80% on LiveCodeBench v6 with strong performance across code generation, debugging, refactoring, and multi-file code understanding tasks.

🌍

Multilingual Support

Trained on 140+ languages with high-quality performance across major world languages, enabling truly global deployment and cross-lingual understanding.

🧠

Advanced Reasoning

Scores 89.2% on AIME 2026 and 84.3% on GPQA Diamond, demonstrating strong mathematical reasoning, scientific analysis, and complex problem-solving abilities.

Benchmark Results

Gemma 4 delivers competitive performance across major academic and industry benchmarks, rivaling much larger proprietary models.

89.2%

AIME 2026

American Invitational Mathematics Examination. Tests advanced mathematical reasoning and problem-solving on competition-level math questions.

80.0%

LiveCodeBench v6

Real-time coding benchmark evaluating code generation, understanding, and debugging across multiple programming languages and difficulty levels.

84.3%

GPQA Diamond

Graduate-level question answering across physics, chemistry, and biology. Tests deep scientific knowledge and multi-step analytical reasoning.

85.2%

MMMLU

Massive Multitask Multilingual Language Understanding. Evaluates broad knowledge and reasoning across 140+ languages and dozens of academic subjects.

Gemma 4 Open Models

Google DeepMind's most capable open-source AI model family. Multimodal understanding across text, image, video, and audio with up to 256K context and Apache 2.0 license.