Google DeepMind has launched Google DeepMind Gemma 4, its latest open model family designed for advanced reasoning and agentic workflows under an Apache 2.0 license. The release introduces four model sizes and expands Google’s push into open AI systems for developers and researchers.
The new lineup includes effective 2B, effective 4B, a 26B Mixture of Experts model, and a 31B dense model. Google said the 31B version currently ranks third among open models on the Arena AI text leaderboard, while the 26B model ranks sixth and outperforms models far larger than itself.
Gemma 4 adds several features aimed at more capable AI applications. The models support native function calls, structured JSON output, and system commands, enabling developers to build autonomous agents that can interact with tools and APIs.
Google DeepMind’s vice president of research said in the official blog post that Gemma 4 delivers an “unprecedented” level of intelligence per parameter. That claim positions the new release as a major step forward for efficient open models.
We just released Gemma 4 — our most intelligent open models to date.
Built from the same world-class research as Gemini 3, Gemma 4 brings breakthrough intelligence directly to your own hardware for advanced reasoning and agentic workflows.
Released under a commercially… pic.twitter.com/W6Tvj9CuHW
— Google (@Google) April 2, 2026Gemma 4 Supports Multimodal input and 140 languages
All Gemma 4 models can process video, image, and text natively. In addition, the E2B and E4B models support native audio input for speech recognition.
Google also said the models support more than 140 languages. Larger versions offer context windows of up to 256K tokens, which can help developers analyse long documents or even entire code repositories in a single prompt.
Google optimised the E2B and E4B models for edge use cases, including mobile and IoT devices. According to the announcement, these versions can run fully offline on phones, Raspberry Pi devices, and NVIDIA Jetson Orin Nano systems with near-zero latency.
The company also said it worked with Qualcomm, MediaTek, and the Pixel team to improve mobile performance. That focus could make Gemma 4 more attractive for developers building lightweight AI tools that need local processing. [Internal link: latest AI model launches] [Internal link: AI models for mobile devices]
Since the first Gemma release, Google said the model family has been downloaded more than 400 million times, helping create a “Gemmaverse” of over 100,000 variants. Users can now access Gemma 4 on Hugging Face, Kaggle, Ollama, and Google AI Studio
