Welcome to Gemma 4 - Google DeepMind Releases Its Most Capable Open Models

Ronni Holmvig Strøm · 2026-04-08

On April 2, 2026, Google DeepMind officially released Gemma 4, a new family of open-weight models that the company is positioning as the strongest openly available AI systems to date.

Unlike previous Gemma releases, Gemma 4 comes with full Apache 2.0 licensing, removing earlier custom restrictions and giving developers complete freedom to use, modify, fine-tune, and commercially deploy the models without additional barriers.

The release includes multiple sizes, ranging from lightweight variants optimized for edge devices all the way up to a powerful 31B-parameter model.

Significant Capability Gains

Early independent evaluations show Gemma 4 delivering substantial improvements over Gemma 3 and closing the gap with leading proprietary models in several key areas:

Strong performance on Arena AI (text) and LiveCodeBench, indicating robust general reasoning and coding abilities. Leading results on AIME 2026 mathematics benchmarks, highlighting advanced quantitative reasoning. Notable gains in multi-step reasoning, structured output generation, and long-context understanding.

Google DeepMind has particularly emphasized that Gemma 4 was purpose-built for agentic workflows. The models were trained with heavy emphasis on tool use, function calling, planning, and execution in real-world environments.

This makes Gemma 4 especially attractive for developers building autonomous agents, research assistants, and production-grade AI applications that require reliable multi-step reasoning.

The Apache 2.0 Shift is A Major Licensing Milestone

One of the most important aspects of the release is the switch to Apache 2.0 licensing. Previous Gemma versions used a more restrictive custom license that limited certain commercial and derivative uses.

By adopting the industry-standard permissive license, Google is directly addressing long-standing developer feedback and opening the door for broader adoption, including in enterprise settings and large-scale fine-tuning efforts.

This change is already generating excitement in the open-source community. With over 400 million downloads of previous Gemma models and a thriving ecosystem of more than 100,000 community variants (the so-called “Gemmaverse”), the new family is expected to accelerate fine-tuning, distillation, quantization, and domain-specific adaptations at an unprecedented pace.