Google releases Gemma 4, an open model family for agentic workflows
Google introduced Gemma 4, an Apache 2.0-licensed set of open-weights models aimed at reasoning and agentic work across devices from phones to workstations.
Quick answer
Instead of calling a hosted chatbot, developers can download model weights and run a Gemma model locally or in their own cloud setup.
What happened
Google published details for Gemma 4 on April 2, 2026, describing four model sizes and positioning the release as open models designed for reasoning and agentic workflows.
Why it matters
Open-weights models give teams the option to run AI on their own hardware and control cost, privacy, and latency. The tradeoff is that developers also take on more responsibility for evaluation and safe use.
Key points
- Released in four sizes to target both edge devices and larger GPUs.
- Licensed under Apache 2.0 for broad developer use.
- Model docs highlight long-context support and function calling for tool-using agents.
What to watch
Watch for independent evaluations, best-practice guidance for fine-tuning, and how quickly popular runtimes and dev tools adopt Gemma 4 defaults.
Key terms
- Open weights
- A model release where parameters are downloadable so others can run it outside the original provider.
- MoE
- Mixture of Experts: a model design that activates only part of the network per request to save compute.
Sources
- Gemma 4: Byte for byte, the most capable open models Google · Primary announcement · Apr 2, 2026 Primary
- Gemma 4 model card Google AI for Developers · Model card · Apr 2, 2026 Primary