Google has unveiled its open-weight AI model family Gemma 4, SiliconANGLE reported on Wednesday.
The company said Gemma 4 is based on the same architecture as Gemini 3. It supports running complex reasoning and autonomous AI agents locally on low-power devices such as workstations and smartphones.
The lineup has 4 models: Effective 2B (E2B), Effective 4B (E4B), a 26B Mixture of Experts (MoE) model and a 31B dense model. E2B and E4B focus on lightweight hardware environments such as Android smartphones or Raspberry Pi. The 26B MoE model is designed to speed up inference by activating only 3.8 billion parameters while maintaining a large-model class knowledge base.
Google said the models natively support function calling and structured JSON output, enabling autonomous agents that link with external tools to carry out multi-step tasks. All 4 models can process image and video inputs. E2B and E4B also support native audio input, providing on-device real-time speech understanding.
Gemma 4 is offered under the Apache 2.0 license. It can be accessed on Google Cloud and is also available as open weights on Hugging Face, Kaggle and Ollama.
Chi-gyu Hwang
delight@d-today.co.kr