Google Gemma 4: The Most Capable Open Models
Byte for byte, the most capable open models — now with Apache 2.0 license.
Multimodal, agent-ready, from edge devices to servers.
Apache 2.0 License — Free for commercial use
Try Gemma 4 Now
Chat with Gemma 4 31B-IT directly in your browser — no setup required.
What is Gemma 4?
Gemma 4 is Google DeepMind's latest open model family built on Gemini 3 technology. It covers 4 model variants from edge devices to workstations, with native multimodal and agent capabilities.
4 Model Variants
E2B (2.3B), E4B (4.5B), 26B MoE, and 31B Dense — covering every use case from mobile to data center.
Native Multimodal
Text, image, video, and audio input support. Smaller models support all four modalities.
Agent-Ready
Built-in function calling, structured JSON output, GUI element detection — no fine-tuning needed.
Apache 2.0 License
First Gemma release under Apache 2.0 — fully free for commercial use, modification, and redistribution.
Why Choose Gemma 4?
Massive performance leap over Gemma 3, with true open-source licensing and full-stack AI capabilities.
4.3x Better Math Reasoning
AIME 2026 score jumped from 20.8% (Gemma 3) to 89.2% (Gemma 4 31B). GPQA Diamond doubled from 42.4% to 84.3%.
20x Coding Improvement
Codeforces ELO surged from 110 to 2150. LiveCodeBench v6 reaches 80% — competitive with top proprietary models.
Run Anywhere — Phone to Server
E2B runs on Raspberry Pi 5 at 133 tokens/sec prefill with under 1.5GB memory. 31B powers server workloads with 256K context.
Key Features of Gemma 4
Everything that makes Gemma 4 the most capable open model family available today.
128K–256K Context Window
E2B/E4B support 128K tokens. 26B MoE and 31B Dense support 256K tokens for long document processing.
Thinking / Reasoning Mode
Enable extended reasoning with enable_thinking=True for complex multi-step problem solving.
Function Calling & JSON Output
Native tool use and structured output generation — ideal for building autonomous AI agents.
Object Detection & GUI Localization
Detect objects in images and precisely locate UI elements — key capability for computer-use agents.
140+ Languages
Multilingual and cross-cultural capabilities with support for over 140 languages out of the box.
Speech Recognition (E2B/E4B)
Smaller models include a 300M-parameter audio encoder for speech recognition and translation.
Gemma 4 by the Numbers
Performance metrics that set Gemma 4 apart from the competition.
89.2% AIME 2026 (Math)
AIME 2026 (Math)
#3 LMArena Open Source
LMArena Open Source
2150 Codeforces ELO
Codeforces ELO
Frequently Asked Questions About Gemma 4
Everything you need to know about Google's latest open model family.
Get Started with Gemma 4 Today
Try the most capable open models — free under Apache 2.0.