Google Releases Gemma 4: Open Models Ranging from 2B to 31B Parameters

Google DeepMind has released Gemma 4, its most intelligent open model family yet, built from Gemini 3 research and technology. The release offers four model sizes optimized for different deployment scenarios.

Model Lineup

Key Capabilities

• Agentic workflows: Native function calling support for autonomous agents
• Multimodal reasoning: Strong audio and visual understanding
• 140 languages: Beyond translation — cultural context awareness
• Fine-tuning support: Compatible with preferred frameworks

Benchmark Performance

Gemma 4 31B achieved an Arena AI score of 1452 (as of April 2, 2026). Notable results include MMMLU at 85.2%, AIME 2026 Mathematics at 89.2%, LiveCodeBench v6 at 80.0%, and GPQA Diamond at 84.3%.

Edge Deployment

The E2B and E4B models can run completely offline with near-zero latency on devices including phones, Raspberry Pi, and Jetson Nano. The 26B and 31B models are optimized for consumer GPUs, enabling frontier intelligence on personal workstations.