Google Releases Gemma 4: An Unprecedented Open-Source AI

Google Releases Gemma 4: An Unprecedented Open-Source AI

Google has recently unveiled Gemma 4, the latest evolution of its series of artificial intelligence models, now available under the Apache 2.0 license. This decision marks a significant turning point, offering truly open access to these models for the first time, without restrictive conditions imposed by Google.

The Apache 2.0 license is well-known in the open-source software field, allowing users to download, modify, redistribute, and even commercialize Gemma 4 without royalties. Since the initial launch of the Gemma series in 2024, these models have been downloaded over 400 million times, generating more than 100,000 variants created by the community. Previous versions were limited by proprietary terms of use, which is no longer the case with Gemma 4.

A Model Designed for Autonomy

Gemma 4 stands out for its ability to operate entirely offline, without requiring a connection to Google's servers. The model is available in four sizes: E2B, E4B, 26B MoE (Mixture of Experts), and 31B Dense. The first two sizes are optimized for mobile and embedded devices, such as Android smartphones, Raspberry Pi, and NVIDIA Jetson, with low memory consumption and minimal latency. The larger models are intended for workstations and local servers, with the 31B Dense capable of running on a single NVIDIA H100 GPU with 80 GB.

In the Arena AI ranking, the 31B Dense ranks third globally for text processing, while the 26B MoE is ranked sixth, a remarkable performance against much larger models.

Advanced Features

In terms of features, Gemma 4 can natively process videos, images, and for the E2B and E4B variants, audio input. The context window capacity reaches 256,000 tokens for the larger models, allowing for the analysis of an entire code repository in a single query. The model supports over 140 languages and integrates function calls as well as structured JSON output, essential for autonomous agent pipelines.

A New Era for Commercial AI

The transition to the Apache 2.0 license unlocks the commercial potential of Gemma. Previous restrictions from Google, which may have deterred companies from adopting these models for critical applications, are now a thing of the past. Clément Delangue, co-founder of Hugging Face, praised this advancement: “The release of Gemma 4 under the Apache 2.0 license is a major milestone. We are excited to support the Gemma 4 family on Hugging Face from day one.”

Google also aims to directly compete with Chinese open-weight models, such as those from Moonshot AI, Alibaba, and Z.AI, which rival giants like GPT-5 or Claude.

For those looking to explore Gemma 4, the model is now available on platforms like Google AI Studio, Hugging Face, Kaggle, and Ollama.