Google has launched Gemma 4, a new generation of its open-weights model family aimed squarely at agentic workflows and on-device inference. The release adds four variants — an Effective 2B (E2B) and Effective 4B (E4B) for edge devices, a 26B Mixture-of-Experts (MoE) and a 31B Dense flagship — all distributed under the Apache 2.0 license.
What changed under the hood
Gemma 4 ships with context windows up to 256K tokens, native vision and audio inputs, native function calling and fluency in more than 140 languages. Google says the 31B Dense model ranks number three on the open-model Arena AI text leaderboard, while the 26B MoE variant lands at number six, with both outperforming several rivals at roughly 20x the parameter count.
Targeted at agents
The release is pitched at developers building agentic systems — multi-step planners, tool-using assistants and on-device copilots — rather than raw chat. Function calling is now a first-class capability, and the smaller E2B and E4B variants are explicitly tuned to run on smartphones, Raspberry Pi-class boards and consumer GPUs, slotting Gemma 4 into Google's broader push for physical and on-device AI.
Distribution and Gemmaverse
Gemma 4 is available through Google AI Studio, AI Edge Gallery and Google Cloud Vertex AI, as well as third-party hubs including Hugging Face, Kaggle and Ollama. Google says the original Gemma family has now been downloaded more than 400 million times, with the community publishing over 100,000 fine-tuned variants — a footprint the company is leaning on as it tries to keep open-weights momentum from drifting toward Chinese model families.
How it stacks up
The release lands in a market where open-weights releases from Mistral, Meta, DeepSeek, Qwen and now Google are competing on a mix of context length, reasoning quality, multimodality and license terms. Gemma 4's combination of Apache 2.0 distribution, native multimodality and a long context window is designed to give enterprises a defensible baseline for building agents that they can host themselves.
Reporting based on coverage from Google, Google Cloud, The Register and TheNextWeb.