Non classé
Gemma 4: Google’s New Generation of Open Source Models
Google launches Gemma 4, its most powerful open source AI models to date. Advanced reasoning, multimodality, 4 formats adapted to every hardware: everything you need to know.
For several years, the battle around open artificial intelligence models (dubbed “open source”) has been raging. On one side, giants like OpenAI or Anthropic rely on proprietary models. On the other, a growing movement pushes to make AI accessible to all, without barriers. It is in this context that Google DeepMind has just launched Gemma 4, its fourth generation of open models.
We just released Gemma 4 — our most intelligent open models to date.
— Google (@Google) April 2, 2026
Built from the same world-class research as Gemini 3, Gemma 4 brings breakthrough intelligence directly to your own hardware for advanced reasoning and agentic workflows.
Released under a commercially… pic.twitter.com/W6Tvj9CuHW
Gemma 4, what exactly is it?
Gemma 4 is a family of open source AI models built from the same research and technologies as Gemini 3, Google’s flagship model. Concretely, this means that Gemma 4 inherits some of Google’s most recent AI advancements. However, it is freely downloadable and usable, including for commercial purposes, thanks to an Apache 2.0 license.
The stated goal is to offer a maximum level of intelligence relative to the number of parameters used. In other words, to do a lot with few resources. On the independent Arena AI ranking, the Gemma 4 31B model ranks third globally among open source models, outperforming models up to 20 times larger.
Four Gemma 4 Model Sizes for All Uses
This is one of Gemma 4’s most differentiating arguments: Google offers four distinct variants, designed for very different hardware environments.
The E2B and E4B models (for “Effective 2B” and “Effective 4B”) are designed for mobile devices and connected objects. They run entirely offline, with near-zero latency, on smartphones, Raspberry Pis, or embedded boards. Furthermore, they natively integrate audio, image, and video processing, ideal for consumer applications.
The 26B and 31B models are aimed at developers and researchers working on more powerful machines. The 26B uses a “Mixture of Experts” (MoE) architecture. It activates only a portion of its parameters at a time (approximately 3.8 billion out of 26), which makes it very fast. The 31B, on the other hand, is a “dense” model that maximizes result quality and offers a solid foundation for fine-tuning (adaptation to specific use cases).
This four-size range is strategic; it allows everyone, whether a student, startup, or large company, to choose the model suited to their hardware and needs, without being forced to invest in costly infrastructure.
Capabilities Designed for Agentic Workflows with Gemma 4
Beyond its raw performance, Gemma 4 introduces important features for modern AI uses. All four models natively support agentic workflows. The AI can call external functions, produce structured data (JSON), and interact with other tools or APIs to accomplish tasks autonomously. This is a major evolution compared to simple chatbots.
Gemma 4 also handles long contexts, up to 256,000 tokens. This is equivalent to hundreds of pages of text processed in a single request. It is capable of understanding and generating content in over 140 languages, and excels in visual tasks such as reading charts or recognizing text in images.
In terms of accessibility, the models are available on Hugging Face, Kaggle, and Ollama, and are compatible with the most popular tools in the ecosystem (vLLM, LM Studio, Keras, etc.).
The Gemma 4 Family from Google: What to Remember?
With Gemma 4, Google sends a strong signal to the entire AI community: cutting-edge artificial intelligence is no longer reserved for companies with massive infrastructures. By offering four sizes adapted to every context, from smartphones to servers, and by relying on Gemini 3 technology, Gemma Google repositions open source as a credible lever against proprietary models.
[1] Gemma 4: Byte by Byte, the Most Performant Open Models
