Skip to content
Google

Gemma 4

Open Source

GoogleReleased on 2026-04-02

Google's most capable open model family. Four sizes optimized for local hardware: E2B and E4B for mobile/edge devices, 26B MoE for speed, 31B Dense for quality. Built on Gemini 3 technology with Apache 2.0 license. Supports 140+ languages, native function calling, agentic workflows, and multimodal input.

83
Overall Score

Voice of the community

sample 156

Gemma 4 brings the first major update to Google's open models in a year.

Ars Technica2026-04-02

Purpose-built for advanced reasoning and agentic workflows, Gemma 4 delivers an unprecedented level of intelligence-per-parameter.

Google Blog2026-04-02

Chain of thoughts on Gemma is concise, helpful and coherent while Qwen does a lot of inner-gaslighting, and also loops a lot on default settings. Visual understanding is very good, and multilingual seems good as well.

r/LocalLLaMA2026-04-04

Core Specs

256K
Context Window
8K
Max Output
ReasoningOpen Sourcetextimagevideoaudio

Pros & Cons

Sentiment75% +20% ·5% −

Pros

  • +Apache 2.0 open-source license (major upgrade from Gemma 3)
  • +Four sizes covering mobile to workstation
  • +26B MoE activates only 3.8B params (fast inference)
  • +Native multimodal: text, image, video, audio
  • +140+ language support
  • +Agentic workflow ready (function calling, JSON output)

Cons

  • Larger models require significant hardware (80GB GPU unquantized)
  • Context window smaller than cloud Gemini (256K vs 1M)
  • Brand new release - ecosystem still developing

Pricing

Input (per 1M tokens)$0.00
Output (per 1M tokens)$0.00
Free trial available
Updated on 2026-04-03

Get Started

1Visit the provider's website
2Create an account
3Start using the model

Benchmarks

arenaElo#3 open model%
mmlu%
humanEval%
arcAgi2%
note31B Dense ranks #3 open model on Arena AI leaderboard (behind GLM-5, Kimi 2.5)%

Reliability

Incidents (30d)0
Self-hosted model. Also available via Google AI Studio.