Google Launches Gemma 4 AI Model

Google has launched the fourth generation of its open-weight Gemma AI model family, marking a significant expansion in capability and accessibility. The new lineup includes four distinct sizes engineered to perform across a spectrum of hardware, from mobile devices to powerful workstations. A pivotal change is the adoption of the permissive Apache 2.0 licence, a major shift from prior versions that opens the door for wider commercial and enterprise adoption. This move was hailed by Hugging Face co-founder Clément Delangue as a substantial milestone for the open-source AI community.

The models share the foundational research and technology powering Google’s advanced Gemini 3 frontier model. Demis Hassabis, CEO of Google DeepMind, asserts these are the leading open models globally for their respective classes. The family is split into two categories: compact edge models and larger, more powerful variants. The Effective 2B (E2B) and Effective 4B (E4B) are designed for on-device operation, running efficiently on smartphones, Raspberry Pi units, and Jetson Nano hardware through collaborations with the Pixel team, Qualcomm, and MediaTek. For more demanding tasks, the 26B Mixture-of-Experts (MoE) and 31B Dense models target offline use on developer hardware and consumer GPUs.

In terms of raw performance, the 31B Dense model already holds an impressive third-place ranking among all open models on the Arena AI text leaderboard, with the 26B MoE variant placing sixth. Google states both larger models outperform competitors up to twenty times their size on this benchmark. Practically, the unquantized weights for the 31B model fit on a single 80GB Nvidia H100 GPU, while quantized versions are compatible with standard consumer graphics cards.

All four models are multimodal, natively processing video and images, and were trained on data spanning over 140 languages. The edge models, E2B and E4B, add native audio input support for speech recognition. They feature a 128K token context window, while the larger 26B and 31B models offer an expanded 256K token context. Key capability enhancements include superior multi-step reasoning, native function-calling, structured JSON output for agentic workflows, and robust offline code generation.

Performance gains are particularly notable for the edge family. According to the Android Developers Blog, the E2B model operates three times faster than the E4B. Overall, this new generation is up to four times faster than previous Gemma versions while consuming up to 60% less battery power. These efficient edge models also form the core of Gemini Nano 4, Google’s forthcoming on-device AI model for Android, scheduled to reach consumer devices later this year.

The Gemma project has seen remarkable traction since its inception, surpassing 400 million downloads and inspiring over 100,000 community-created variants, a metric Google cites as proof of widespread developer engagement. The new Gemma 4 models are available now on platforms like Hugging Face, Kaggle, and Ollama. The 31B and 26B models are accessible through Google AI Studio, with the edge models offered via the AI Edge Gallery. The strategic shift to Apache 2.0 licensing is perhaps the most impactful commercial aspect of this release, explicitly removing prior restrictions to enable a far broader range of real-world, production applications.