Gemma 4: The Most Capable Open Models Yet

Google DeepMind announces Gemma 4, a family of state-of-the-art open-source models released under Apache 2.0 license.

Model Variants

  • Gemma 4 E2B (Effective 2B): Edge-optimized, audio + text + image support
  • Gemma 4 E4B (Effective 4B): Edge-optimized, audio + text + image support
  • Gemma 4 26B MoE: Mixture of Experts variant, high performance
  • Gemma 4 31B Dense: Flagship dense model, #3 on Arena AI benchmarks

Key Features

  • Apache 2.0 License: Fully permissive, commercial-friendly
  • Multimodal: Text, image, video, audio (on edge models)
  • Agentic AI Focus: Purpose-built for advanced reasoning and agent workflows
  • Context Window: Edge models 128K tokens, larger models 256K tokens
  • Language Support: 140+ languages
  • Performance: 31B variant #3 on Arena AI text leaderboard, 26B variant #6

Deployment Options

  • Smartphones and edge devices
  • Raspberry Pi and IoT
  • Local on-device deployment
  • Google Cloud integration
  • HuggingFace ecosystem

Community Adoption

  • Prior Gemma versions (1-3): 400M+ downloads
  • 100K+ community variants in HuggingFace ecosystem
  • Extended support for fine-tuning and RAG applications

Release Timeline

  • Private release: March 31, 2026
  • Public release: April 2, 2026

Strategic Positioning

Gemma 4 represents Google’s dual-track LLM strategy:

  • Proprietary: Gemini (closed, frontier performance)
  • Open-source: Gemma 4 (developer-friendly, local deployment, Apache 2.0)