Contradiction

Slug collision: [[gemma-4]] exists as both concept and entity. Prefer concept page (wiki/concepts/gemma-4.md) for model-family details; entity stub is thin/incomplete.

This page may contain stale information. Last updated: 2026-04-22

Definition

Gemma 4 is Google DeepMind’s latest family of open-source large language models released under the Apache 2.0 license, designed specifically for advanced reasoning and agentic AI workflows with multimodal capabilities (text, image, video, audio).

Key Characteristics

  • License: Apache 2.0 (fully permissive, commercial-friendly)
  • Model Variants: 4 sizes (E2B, E4B, 26B MoE, 31B Dense)
  • Multimodal: Native support for text, image, video, audio
  • Context Windows: 128K (edge) to 256K (dense) tokens
  • Languages: 140+ languages supported
  • Performance: 31B variant #3 on Arena AI leaderboard, 26B variant #6
  • Target: On-device AI, edge deployment, agentic workflows

Model Variants

ModelSizeTypeContextSpecialization
E2B2B (effective)Dense128KEdge devices, smartphones
E4B4B (effective)Dense128KEdge devices, IoT
26B MoE26B (mixture)Mixture of Experts256KBalanced performance
31B Dense31BDense256KFlagship, highest performance

Performance Highlights

Arena AI Leaderboard

  • 31B variant: #3 ranked open-source model (world’s leading)
  • 26B variant: #6 ranked open-source model
  • Competitive with proprietary frontier models in reasoning tasks

Agentic Capabilities

  • Advanced reasoning for multi-step tasks
  • Tool use and function calling
  • Planning and decomposition
  • Collaborative workflows

Multimodal Support

  • Visual: Image understanding, OCR, chart comprehension, variable resolution
  • Video: Long-form video understanding, video question answering
  • Audio: Speech recognition, audio understanding (E2B, E4B)
  • Text: All variants, 140+ languages

Deployment Options

  • Local: Offline on-device deployment
  • Edge: Smartphones, tablets, IoT devices
  • Computing: Raspberry Pi to high-end GPU clusters
  • Cloud: Google Cloud integration, HuggingFace ecosystem
  • Ecosystem: Full HuggingFace support for fine-tuning, RAG

Strategic Position

Part of google’s dual-track LLM strategy:

  • Proprietary (Gemini): Closed-source, frontier performance
  • Open-source (Gemma 4): Developer-friendly, local deployment, commercial freedom

Differentiates from openai (closed-source) and mistral (similar positioning but smaller models).

Community Adoption

  • Prior versions: 400M+ downloads (Gemma 1-3)
  • Community variants: 100K+ on HuggingFace
  • Research integration: Academia, startups, enterprises
  • Extended support: Fine-tuning, RAG, custom applications

Release Timeline

  • Private release: March 31, 2026
  • Public release: April 2, 2026
  • Adoption: Rapid community integration

Technical Innovation

Agentic Focus

Explicitly designed for:

  • Complex reasoning
  • Tool use integration
  • Sequential decision-making
  • Autonomous task execution

Size Efficiency

  • E2B/E4B: Run on edge devices with full multimodal capabilities
  • 26B MoE: Efficient inference through sparse activation
  • 31B Dense: High performance while maintaining reasonable inference cost

Sources