Robots Atlas>ROBOTS ATLAS
Gemma 4
AI Modelsโ€บGemma

Gemma 4

Family: Gemma
Open (Apache 2.0) family of multimodal AI models from Google DeepMind (E2B/E4B/26B A4B/31B). Supports text, image, audio, and video. Native function calling.
โœ“ Activeโœ“ Public accessโš– Open sourceโ˜… FeaturedLLMMultimodalTool-using model๐Ÿ“ Gemma
Context window
256K
tokens
Parameters
25.2B
parameters
๐ŸขGoogle DeepMindProducer
Access:APIDownloadHostedDeployment:๐Ÿ’ป Localโ˜ Cloud๐Ÿ“ฑ On-device

Overview

Gemma 4 is a family of open multimodal AI models from Google DeepMind, released under the Apache 2.0 license. The models support text, image, audio (E2B/E4B), and video input, and generate text, code, and structured data.

**Model variants**

Gemma 4 is available in four sizes: E2B and E4B (mobile and edge devices, 128K token context window), 26B A4B (Mixture of Experts, consumer GPU, 256K context window), and 31B (dense, workstation-class GPU, 256K context window).

**Key capabilities**

Built-in reasoning mode (Thinking) โ€” models feature native chain-of-thought

Native function calling (native function-calling support)

Native system prompt support (native system role)

Hybrid attention mechanism: local sliding window + global attention (final layer is always global)

The 26B A4B and E4B models operate at speeds comparable to 4B models due to the MoE architecture

Classification
LLMMultimodalTool-using model
Family: Gemma
Access & deployment
APIDownloadHosted
LocalCloudOn-device
Weights: Open source
Key parameters
๐Ÿ“ Context: 256K
๐Ÿงฉ Parameters: 25.2B
โœ“ Toolsย ยทย โœ“ Fine-tuning
๐Ÿ“ฅ Input: text, image, audio, video
Platforms
Vertex AIHugging Face Hub

Technical specification

Context window
256K
tokens
Parameters
25.2B
parameters
License
Apache 2.0
Hardware requirements
E2B/E4B: mobile and edge devices (phones, tablets, IoT); 26B A4B (MoE): consumer GPU or workstation; 31B: workstation-class GPU. The E2B/E4B variants are designed to run on devices without cloud connectivity.
Features:โœ“ Tool useโœ“ Fine-tuning
Modalities
โฌ‡ Input
textimageaudiovideo
โฌ† Output
textcodestructured_data

Capabilities and applications

Native model capabilities
Coding
Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.
Category: coding
Multilingual
Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.
Category: language
Multi-step reasoning
Carrying out multi-step chains of reasoning across long, complex tasks.
Category: reasoning
Long context
Support for large context windows โ€” tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.
Category: language
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Image understanding
Analysing and interpreting the content of images.
Category: vision
Audio understanding
Category: audio
Multimodal understanding
Category: multimodal
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning
Video Understanding
Category: video
Function Calling
Category: planning
Interleaved Multimodal Input
Category: reasoning

Benchmark results

5 benchmarks
MMLU Pro
accuracy ยท instruction-tuned (Gemma 4 31B IT)
85.2%
๐Ÿ“… 31 Mar 2026๐Ÿ“„ Gemma 4 model card | Google AI for Developers
Source: ai.google.dev/gemma/docs/core/model_card_4. Score for Gemma 4 31B (IT). MMLU Pro is harder than standard MMLU.
GPQA
accuracy ยท Instruction-tuned variant (Gemma 4 31B IT).
84.3%
๐Ÿ“… 31 Mar 2026๐Ÿ“„ Gemma 4 model card | Google AI for Developers
Source: ai.google.dev/gemma/docs/core/model_card_4. Result for Gemma 4 31B (IT). Diamond subset of GPQA.
LiveCodeBench v6
accuracy ยท instruction-tuned (Gemma 4 31B IT)
80.0%
๐Ÿ“… 31 Mar 2026๐Ÿ“„ Gemma 4 model card | Google AI for Developers
Source: ai.google.dev/gemma/docs/core/model_card_4. Result for Gemma 4 31B (IT). LiveCodeBench v6 coding benchmark.
AIME 2026 (no tools)
accuracy ยท Instruction-tuned variant (Gemma 4 31B IT), no tools enabled.
89.2%
๐Ÿ“… 31 Mar 2026๐Ÿ“„ Gemma 4 model card | Google AI for Developers
Source: ai.google.dev/gemma/docs/core/model_card_4. Score for Gemma 4 31B (IT). American Invitational Mathematics Examination 2026.
MMMU Pro (Vision)
accuracy ยท Instruction-tuned variant (Gemma 4 31B IT), vision-capable model.
76.9%
๐Ÿ“… 31 Mar 2026๐Ÿ“„ Gemma 4 model card | Google AI for Developers
Source: ai.google.dev/gemma/docs/core/model_card_4. Score for Gemma 4 31B (IT). MMMU Pro is an extended, more challenging version of MMMU.

Pricing

Technical architecture

Core Architecture
MOMoE
Model Form
MLMultimodal LLM
Training Techniques
ITInstruction TuningRLRLHFP/PEFT / LoRA

Deployment and security

๐Ÿ”’ Security / Enterprise
โœ“ Verified enterprise information

Gemma 4 model card includes safety evaluation results. As an open-source model, deployment responsibility lies with the user. Documentation on responsible AI use is available.

Gemma 4 is an open-source model (Apache 2.0). Production deployments require independent risk assessment. Google publishes a model card with safety evaluation results.
Updated: 5 Apr 2026โ†— Security documentation

Sources and related pages