Google open model family

Gemma 4

Google Gemma 4 is an open-weight multimodal model family built for advanced reasoning, coding, agentic workflows, and efficient local deployment from phones to workstations.

256K
maximum context window
5 sizes
from edge to workstation
Apache 2.0
open-source license
Gemma 4 benchmark

Official Gemma 4 benchmark results

Google reports strong results for the 31B instruction-tuned model across reasoning, coding, agents, vision, and long-context evaluation.

MMLU Pro

85.2%

Advanced knowledge and reasoning score for Gemma 4 31B.

AIME 2026

89.2%

Competition mathematics without tools for Gemma 4 31B.

LiveCodeBench v6

80.0%

Code-generation benchmark score for Gemma 4 31B.

GPQA Diamond

84.3%

Graduate-level science reasoning score for Gemma 4 31B.

Tau2

76.9%

Average agentic task score across three domains for Gemma 4 31B.

MMMU Pro

76.9%

Multimodal reasoning score for Gemma 4 31B.

Scores come from Google’s Gemma 4 model card. Vendor-reported benchmark settings differ, so evaluate the model on your own prompts, hardware, and workflows.

Use cases

Where Gemma 4 fits

The right Gemma 4 variant depends on the quality, modality, privacy, latency, and hardware requirements of your workflow.

Local coding assistants

Run code generation, completion, correction, and repository analysis on your own workstation.

Private document analysis

Process long documents, images, charts, and OCR tasks locally when data control matters.

On-device multimodal apps

Use E2B or E4B for efficient image, audio, and text experiences on edge hardware.

Tool-using agents

Build agent workflows with function calling, structured output, system instructions, and configurable thinking.

Gemma 4 models

Choose a Gemma 4 model for your hardware

The family scales from efficient on-device models to larger dense and Mixture-of-Experts models for laptops, consumer GPUs, and workstations.

ModelArchitectureContextModalitiesBest for
Gemma 4 E2B2.3B effective dense128KText, image, audioPhones and edge devices
Gemma 4 E4B4.5B effective dense128KText, image, audioEdge devices and lightweight local AI
Gemma 4 12B11.95B unified dense256KText, image, audioMultimodal agents on laptops
Gemma 4 26B A4B25.2B MoE, 3.8B active256KText, imageFast workstation inference
Gemma 4 31B30.7B dense256KText, imageMaximum quality and fine-tuning
Architecture

Efficient by design across five model sizes

Gemma 4 uses several architecture strategies to balance reasoning quality, multimodal capability, speed, and memory use.

01

Hybrid attention

Local sliding-window attention is interleaved with global attention to reduce memory use while preserving long-context awareness.

02

Dense and MoE options

Choose dense models for straightforward quality scaling or the 26B A4B MoE model for speed with only 3.8B active parameters.

03

Unified 12B multimodality

Gemma 4 12B projects image patches and raw audio into one decoder-only transformer without separate multimodal encoders.

04

Quantization-aware checkpoints

Google provides QAT checkpoints designed to reduce memory requirements and improve practical local deployment.

Gemma 4 FAQs

Run Google Gemma 4 AI in a hosted MyClaw agent

Use MyClaw to run always-on AI agents with files, tools, browser-ready execution, and managed infrastructure.

Run Gemma 4 in MyClaw