Google Launches Gemma 4: Bringing “Gemini 3” Intelligence To The Open-Source Community

MOUNTAIN VIEW — Google DeepMind has officially released Gemma 4, the latest generation of its open-weight large language models. Building on the technological foundations of the proprietary Gemini 3 Pro released late last year, Gemma 4 is designed to bring “frontier-level” intelligence to developers, researchers, and enterprises for local and edge computing.

1. The Gemma 4 Family: From Micro-Devices to High-End Servers
   Gemma 4 is being released in four distinct sizes to cater to different hardware constraints and use cases:

Effective 2B (E2B) & Effective 4B (E4B):

Target: Smartphones, IoT devices, and edge hardware (Raspberry Pi, Nvidia Jetson Orin Nano).

Optimization: Specifically tuned to preserve RAM and battery life while enabling multimodal inference on power-constrained systems.

26B Mixture of Experts (MoE) & 31B Dense:

Target: High-end workstations and servers (Nvidia RTX GPUs, H100, DGX Spark).

Performance: Designed for advanced reasoning and agentic AI (autonomous workflows). The 31B model currently ranks #3 globally on the Arena AI leaderboard, outperforming models many times its size.

2. Key Technical Breakthroughs
   Gemma 4 marks a significant leap over its predecessors in terms of flexibility and raw intelligence:

Multilingual Mastery: Natively trained on over 140 languages, making it one of the most linguistically diverse open models available.

Massive Context Window: Supports a standard 128K context window, with some variants scaling up to 256K, allowing for the processing of entire codebases or long documents.

Apache 2.0 License: In a major shift for Google, these models are released under a highly permissive license, allowing for unrestricted commercial use, modification, and redistribution.

Privacy & Cost: By running locally (offline), Gemma 4 eliminates cloud latency, reduces API costs, and ensures that sensitive data never leaves the user’s device.

3. Gemma vs. Gemini: What’s the Difference?
   While both share the same research DNA, they serve different purposes:

Gemini: A closed, “chatbot-esque” implementation managed by Google (like the one you are talking to now).

Gemma: An AI processing engine. It is a “model-only” release meant to be integrated into other applications, fine-tuned by developers for specific tasks (like coding assistants or medical research tools), and run on private hardware.

Gemma 4 Global Performance Rankings (Arena AI)

Industry Impact
Clement Farabet, VP of Research at Google DeepMind, noted that the “Gemmaverse” has already seen over 400 million downloads since its inception. Partnering with Nvidia, Qualcomm, and MediaTek, Google is positioning Gemma 4 to be the backbone of the next generation of “Agentic AI”—software that doesn’t just answer questions but can autonomously execute complex tasks across different apps and platforms.