Google launches Gemma 4 open AI models for developers

Google on April 2, 2026 introduced Gemma 4, a family of open AI models offered in four variants: Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts (MoE) and 31B Dense. The company described Gemma 4 as its “most capable open model” to date. Google DeepMind CEO Demis Hassabis confirmed the development in a social post.

Google released the models to address different performance and hardware needs. The lineup is intended to let developers run advanced AI tasks on personal workstations, edge devices and cloud servers rather than only in large data centers.

E2B has roughly 2 billion parameters and E4B has about 4 billion. The 26B MoE uses a Mixture of Experts architecture that activates only a subset of the model for each request, lowering computation compared with dense models that run all parameters for every request. The 31B Dense model is the largest in the family; Google reported that it ranks among the top-performing open models on common industry benchmarks.

Gemma 4 supports advanced reasoning, agentic workflows, coding assistance, complex mathematical problem solving and more than 140 languages. Google highlighted the models’ ability to generate high-quality code and to address technical problems, positioning them as potential local AI coding assistants for developers working outside large cloud environments.

Google emphasized efficiency as a design goal. Smaller and optimized variants let developers trade compute and latency for capability. In general, models with more parameters may deliver higher performance but require greater computing resources.

Technical specifications and capability details appear in Google’s blog post and in the company leadership social post. The company released the models to the developer community in the four announced variants.