Google unveils 4 AI that can run on your smartphone or computer, without Internet

Google has just made a big move in the artificial intelligence sector by introducing Gemma 4. Important note: it's not one, but rather four AI models that were presented in a video by the French Olivier Lacombe, software engineer at Google DeepMind. 

Distributed under the Apache 2.0 license. and based on the architecture of Gemini 3, Gemma 4 is therefore available in four versions: Gemma 4 E2B (lightweight and fast, designed to run on smartphones), Gemma 4 E4B (also designed for smartphones, but capable of more complex reasoning and handling heavier tasks), Gemma 4 26B (intended for laptops due to its higher performance), and finally Gemma 4 31B Dense (the heaviest and most powerful model, designed for workstations and servers).

We just released Gemma 4 — our most intelligent open models to date.

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

— Google (@Google) April 2, 2026

Each of these models is therefore suited to performance needs, whether on smartphone, laptop or desktop computer.

An open source model

The major difference between Gemma 4 and Gemini 3 is its open source nature. Specifically, the Gemma 4 models, under Apache 2.0 license, can be downloaded locally, modified, and even commercialized. In other words, the different models are open to developers. Therefore, the open-source nature of Gemma 4 allows those who download one of the four models to optimize their tools and significantly enhance privacy, as each can run offline.

For Google, Gemma 4 allows the American company to provide a guarantee of transparency to its ecosystem, but more importantly, it positions itself as a strong competitor against certain Chinese giants who have released their AI (partially) as open source, such as DeepSeek or Qwen (Alibaba). For yes, the Gemma 4 models rival, or even surpass, some models from DeepSeek and Qwen.

Impressive performance

The features of Gemma 4 presented by Google are impressive: trained on over 140 languages, context window up to 128K or 256K (depending on the model), native functions for deploying autonomous agents, offline code generation, integrated reasoning mode that allows the model to think step by step before responding, native multimodal support (understands and processes different types of content, such as text, videos, photos, etc.).

As we can see from the graph below, the 31B model, which is the most powerful in the Gemma 4 family, ranks third among open models on the Arena AI benchmark. Thus, it competes with Qwen 3.5-397B-A17B and surpasses Deepseek-v3.2-exp-thinking. 

With Gemma 4, Google thus responds to the impressive advances made by China in the field of open-source AI, while reinforcing its leadership position in the industry.

👉🏻 Follow real-time tech news : add 01net to your Google sources, and subscribe to our WhatsApp channel.

Source : Google