Google DeepMind Releases Multimodal Gemma 4 Under Permissive Apache 2.0 License

Google DeepMind has launched Gemma 4, a major open-source AI release under the permissive Apache 2.0 license, marking a significant shift in its open model strategy. The release includes four distinct multimodal models designed for different use cases: the Efficient 2B and 4B models are optimized for phones and edge devices with native audio input support, while the larger 26B Mixture of Experts and 31B Dense models target high-performance applications. The flagship 31B model has already secured the third global ranking among all open models on Arena AI with an Elo score of 1452, achieving 89.2% on the AIME 2026 benchmark and 80.0% on LiveCodeBench v6. All models support text, images, and video, with the larger variants featuring a 256,000-token context window.

Day-one deployment support is extensive, covering Hugging Face, Ollama, vLLM, llama.cpp, MLX, LM Studio, NVIDIA NIM, and Android Studio, making integration straightforward for developers and researchers. Users can immediately run the model via `ollama run gemma4:27b`. This release represents Google DeepMind's most accessible open model family to date, providing powerful, commercially usable AI tools that could accelerate innovation in edge computing, multimodal applications, and agent development, while fostering greater competition in the open-source AI landscape.