Gemma 4: Google Releases Open Model with 31 Billion Parameters

Gemma 4: Google Releases Open Model with 31 Billion Parameters

Google DeepMind has released the fourth generation of its open language models — Gemma 4. The lineup includes four variants: the compact E2B and E4B for mobile devices, a 26B model with a Mixture of Experts architecture, and the flagship 31B Dense model. All versions are available under the Apache 2.0 license and have already been published on Hugging Face.

What is Mixture of Experts Architecture

A key innovation is the 26B MoE (Mixture of Experts) model. With a total of 26 billion parameters, it activates only 3.8 billion per request. This allows it to retain the depth of knowledge of a large model while operating with lower memory requirements and generating responses faster.

How it works: the model is divided into specialized blocks called "experts." Only the necessary set is activated for each task. For example, specific experts are engaged for programming, while others are used for translation. This is the first time Google has applied this architecture to the open Gemma lineup.

Benchmark Results

The flagship 31B Dense model has taken third place among all open models in the Arena AI Text ranking. According to researchers Clément Farabet and Olivier Lacombe, the team aimed to maximize "intelligence per parameter," and the result outperformed models twice its size.

For comparison, the 31B version competes with closed models of 60–70 billion parameters in terms of reasoning and programming quality. At the same time, it fits on a single GPU and operates significantly faster.

Multimodality and Language Support

All four Gemma 4 models can process not only text but also images and video. The smaller E2B and E4B versions also support audio input, enabling on-device speech recognition without sending data to a server.

The context window is up to 128,000 tokens for the compact models and up to 256,000 for the larger ones. More than 140 languages are supported, including Russian. The models are natively designed for agentic scenarios, featuring built-in support for function calling and structured JSON output.

Where to Run It

Gemma 4 is available on several platforms:

  • Hugging Face — for developers and researchers
  • Kaggle — with ready-to-use notebooks for experimentation
  • Ollama and LM Studio — for local execution on your computer
  • Docker — for deployment in the cloud or on your own servers

The smaller E2B and E4B versions can run on Android smartphones, Raspberry Pi single-board computers, and Jetson Nano embedded systems. This makes them suitable for offline applications such as voice assistants, translators, and recognition systems.

Who Is Gemma 4 For

The open Apache 2.0 license removes most commercial restrictions. Models can be used in products without royalties to Google, and the architecture can be modified or fine-tuned on your own data.

Primary use cases include:

  • Developing low-latency chatbots and assistants
  • On-device document processing and data analysis
  • Prototyping AI features without dependency on cloud APIs
  • Tasks with requirements for digital sovereignty and privacy

The flagship 31B model fits on a single GPU, making it accessible to small teams and startups that cannot afford the infrastructure required for models with hundreds of billions of parameters.

Try It in GEMERA AI

All models in the Gemma family, including the new fourth version, are available via GEMERA AI — in our Telegram bot and web dashboard. You can test different sizes, compare speed and response quality, and evaluate multimodal capabilities. An API with flexible settings for context and generation parameters is available for developers.