AI Models
← Back to all models

Gemma 4

GoogleLarge Language ModelOpen Source

Google's current open-weight model family, released March 2026 under Apache 2.0. Sizes E2B, E4B, 26B, and 31B for on-device and self-hosted deployments; the Gemma family passed 1 billion downloads in August 2026.

Abilities

Text GenerationCode GenerationVisionMultilingualOn-deviceFine-tuning

Use Cases

On-deviceChatbotCoding AssistantEducationAutomation

Available in Tools

Ollamahttps://ollama.com
Hugging Facehttps://huggingface.co/google
Google AI Studiohttps://aistudio.google.com

Availability

Hugging Facehttps://huggingface.co/google
Ollama
Gemini API (free tier)https://ai.google.dev
Vertex AI

How to Use

Download open weights from Hugging Face or run locally via Ollama. The E2B/E4B efficient variants target on-device use; 26B and 31B suit self-hosted servers. Also served free-tier-only in the Gemini API.

Pros

  • Apache 2.0 license — fully permissive for commercial use, unlike the older Gemma license
  • Efficient E2B/E4B variants run on phones and edge devices
  • Derived from Gemini research — strong quality for its size class
  • Massive ecosystem — the Gemma family passed 1B downloads in August 2026

Cons

  • Far below frontier models on complex reasoning and coding
  • Largest size is 31B — no large-scale MoE variant like Llama 4 or Qwen
  • Gemini API access is free-tier only — production API serving requires self-hosting or Vertex AI

What to Use It For

star

Perfect For

On-device AI applications

E2B/E4B variants are designed for phones and edge hardware with no cloud dependency

Commercial self-hosted deployments

Apache 2.0 removes the legal friction of the previous Gemma license

thumb_up

Good For

Fine-tuning for domain tasks

Open weights in practical sizes make custom training affordable

Privacy-sensitive workloads

Runs fully offline — no data leaves your infrastructure

warning

Not Recommended

Frontier-level reasoning

Small open models cannot match frontier reasoning depth

Try instead: Claude Opus 5, Gemini 3.8 Flash

Very long context processing

Context is far below the 1M+ windows of hosted frontier models

Try instead: Gemini 3.8 Flash, Llama 4 Scout

block

Do Not Use For

Image or video generation

Text and vision-understanding model only

Try instead: FLUX.1, Stable Diffusion 3.5

Technical Details

PricingFree (open weights, Apache 2.0)
ParametersE2B / E4B / 26B / 31B
detail.contextWindow128K tokens
LicenseApache 2.0

Links

language Website description API Documentation code Repository