Google launched Gemma 4: the open source model that beats rivals 20 times its size

Google DeepMind just dropped what many consider a watershed moment for open source models: Gemma 4, built on the same technology as Gemini 3 (Google's most powerful AI), but available for anyone to download, modify, and use — even commercially.

What sizes are available?

Gemma 4 comes in 4 versions designed for different use cases:

• E2B (Effective 2B): The smallest. Runs on your Android phone or laptop without a dedicated GPU. Ideal for apps that need local AI without relying on internet.
• E4B (Effective 4B): Slightly more powerful, also designed for devices. Perfect for on-device assistants that need more reasoning capability.
• 26B MoE (Mixture of Experts): A smart model that doesn't use all its parameters at once — it only activates the "experts" needed for each task. Ranks #6 globally among open models.
• 31B Dense: The powerhouse. #3 globally on the Arena AI leaderboard among open source models, outperforming models with 20 times more parameters. It's like a compact car beating a truck in a race.

What can it do?

Gemma 4 isn't just for chatting. Here are its main capabilities:

• Advanced reasoning: Can solve complex problems step by step, like a tutor explaining math by breaking down each part.
• Autonomous agents: Supports function-calling and JSON output, meaning it can execute real actions — search for information, call APIs, fill out forms — not just generate text.
• Code: Generates and understands code in multiple languages, ideal for automating development tasks.
• Vision and audio: Understands images, video, charts, OCR (reading text in photos), and even audio. You can show it a photo of a receipt and have it organize it into a table.

Why does open source matter?

Under the Apache 2.0 license, Gemma 4 is completely free. This means a startup in Buenos Aires, a researcher in Nigeria, or a freelance developer in Vietnam can use the same model Google uses internally — without paying licenses, without asking permission, without usage limits. It's as if Toyota released the blueprints for their best engine so any workshop in the world could build it.

Real use cases already happening:

• Yale University used Gemma for their Cell2Sentence-Scale project, discovering new pathways for cancer therapy. Literally, open source AI helping find medical treatments.
• INSAIT created BgGPT, the first Bulgarian language model, using Gemma as a base. This shows how an open model can democratize AI for languages that big tech ignores.

Gemma community numbers:

Since the first Gemma launch, the family has accumulated over 400 million downloads and more than 100,000 community-created variants (what Google calls the "Gemmaverse"). It's the most active open source AI ecosystem in the world.

Gemma 4 is now available on Google AI Studio and Hugging Face for you to try today.