The bottom line: Google DeepMind introduces Gemma 4 – high-performance open-source models in four sizes (2B to 31B parameters). The 31B model ranks 3rd globally, outperforming models 20x its size and runs efficiently on edge devices through to professional GPUs.
Google DeepMind has unveiled Gemma 4 – a new generation of powerful open-source language models designed for agent workflows and complex reasoning. With four different sizes, Gemma 4 offers unprecedented intelligence per parameter and ranks among the world’s top models in its larger variants.
Google DeepMind presents Gemma 4, a breakthrough generation of open language models specifically designed for advanced logical reasoning and agentic workflows. The new family builds on the same research and technology as Gemini 3 and sets new benchmarks in efficiency per parameter.
The Gemma 4 portfolio comprises four variants: Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts (MoE), and 31B Dense. The performance of the larger models is particularly noteworthy: the 31B model currently ranks 3rd among the world’s best open-source models on the Arena AI leaderboard, with the 26B model in 6th place. This is especially impressive given that the 26B model outperforms competitors that are 20 times larger.
The smaller models (E2B and E4B) were specifically optimized for edge computing and prioritize multimodality, low latency, and seamless ecosystem integration. This enables efficient execution on billions of Android devices worldwide, laptop GPUs, and workstations.
The entire Gemma 4 portfolio is available under the Apache 2.0 license. Since the introduction of the first generation, Gemma models have been downloaded over 400 million times and inspire a vibrant ecosystem with more than 100,000 variants. Projects such as Bulgarian-language BgGPT and the collaboration with Yale University on Cell2Sentence-Scale demonstrate the innovation potential of this platform.