Based on the same research as Gemini 3, the new family consists of a 2B edge model running on Raspberry Pi to a 31B dense model, currently ranked third on the Arena AI open model leaderboard. The Apache 2.0 license is a significant change from previous Gemma releases.
Google released Gemma 4the latest generation of a four-size, lightweight model family designed to cover everything from in-device rendering on smartphones to workstation-class deployments.
Models are made from here The same research and technology that underpins Gemini 3is Google’s own boundary model and is released under the Apache 2.0 license, with more permissive terms than previous Gemma generations and a change described by Hugging Face co-founder Clément Delangue “It’s a big milestone.”
Demis Hassabis, CEO of Google DeepMind, called the new models “world’s best outdoor models for their respective sizes.”
The four variants are the Edge 2B (E2B) and Edge 4B (E4B) models designed to run on-device in phones, Raspberry Pi and Jetson Nano hardware developed in collaboration with the Pixel team, Qualcomm and MediaTek; and 26B Mixture-of-Experts (MN) and 31B Dense models, aimed at offline use on developer hardware and consumer GPUs.
The 31B Dense model currently ranks third among all open models on the Arena AI text leaderboard; 26B TN is in sixth place. Google claims that both large models compete with up to 20 times their size on this metric.
The 31B’s unweighted weights correspond to a single 80GB Nvidia H100 GPU; Quantized versions run on consumer hardware.
All four models are multimodal, natively process video and images, and are taught in more than 140 languages. The E2B and E4B models additionally support native audio input for speech recognition. Context windows are 128K characters for edge models and 256K characters for the two larger variants.
In terms of capabilities, Google highlights multi-step reasoning improvements, native function calling and structured JSON output for agent workflows, and offline code generation. In terms of performance, the Android Developers Blog notes that the E2B model runs three times faster than the E4B, while the edge family is up to four times faster than previous Gemma versions and uses up to 60% less battery.
The E2B and E4B models also form the basis of Gemini Nano 4, Google’s next-generation in-device model for Android, which will arrive in consumer devices later this year.
Gemma has amassed over 400 million downloads and over 100,000 community-created variants since its initial release, a number Google cites as evidence of developer adoption at scale.
Gemma 4 is available immediately on Hugging Face, Kaggle and Ollama, models 31B and 26B are available through Google AI Studio, and edge models are available through AI Edge Gallery.
The Apache 2.0 licensing decision is the most consequential commercial signal at launch: it removes restrictions that prevented some enterprise and commercial deployments under the previous Gemma terms, and opens up the ecosystem to a wider range of production use cases.
