Google Releases Gemma 4: Open Models That Outperform Rivals 20x Their Size
Google launches Gemma 4, a family of open models ranging from 2B to 31B parameters that deliver unprecedented intelligence-per-parameter, outperforming models 20 times their size on key benchmarks.
Google has released Gemma 4, its most capable open model family to date, marking a major leap in the open-source AI landscape. The new family includes four variants — E2B (2B parameters), E4B (4B parameters), a 26B Mixture of Experts model, and a 31B Dense model — each targeting different deployment scenarios from edge devices to cloud infrastructure.
What sets Gemma 4 apart is its remarkable efficiency. The 31B Dense model consistently outperforms closed-source models with parameter counts 10 to 20 times larger, according to Google's benchmarks. With over 400 million cumulative downloads across the Gemma series, the new release builds on a proven track record of developer adoption and real-world deployment. All models are released under the Apache 2.0 license, giving developers broad rights to use, modify, and distribute them freely.
Gemma 4 introduces native multimodal capabilities across all model sizes, accepting both text and image inputs, with the edge-focused variants additionally supporting audio. The models feature long context windows of up to 256K tokens and are purpose-built for advanced reasoning and agentic workflows. Support for over 140 languages makes Gemma 4 one of the most accessible open model families available. Google has made the models available through Hugging Face, Google AI Studio, Vertex AI, and Google Cloud.
The release signals Google's deepening commitment to open AI infrastructure at a time when open-weight models are increasingly competitive with proprietary alternatives. Developers can run Gemma 4's smallest models on devices as modest as a Raspberry Pi or smartphone, bringing frontier-class AI to entirely offline deployments. With this release, Google is positioning Gemma 4 not just as a research artifact but as production-ready infrastructure for the next generation of AI applications.
