Google ships Gemma 4 under Apache 2.0 and it outperforms models 20 times its size

Google released Gemma 4 on April 2, its first major update to the open model family in over a year. The release includes four model sizes, a license change that the open-source community has been asking for since day one, and performance numbers that punch well above their weight class.

Four sizes, two target audiences

The Gemma 4 family splits cleanly between edge and workstation deployment:

• E2B (Effective 2B) and E4B (Effective 4B) - built for phones, Raspberry Pi, Jetson Nano, and similar hardware. Google worked with Qualcomm and MediaTek on these. They support image, text, and audio input with near-zero latency and 128k context windows.
• 26B MoE (Mixture of Experts) - 26 billion total parameters but only activates 3.8B during inference. Runs unquantized on a single H100. 256k context.
• 31B Dense - the flagship. Full dense model, also fits on a single H100 unquantized. 256k context.

All four models share the same underlying architecture as Gemini 3, Google closed-source model family. That means you get the reasoning improvements, math capabilities, and instruction-following that Gemini 3 brought, but with the freedom to run it on your own hardware.

The Apache 2.0 switch

This might be the most important part of the release. Previous Gemma versions shipped under a custom Google license that developers found restrictive. The Gemma 3 license had a prohibited-use policy Google could update unilaterally, required downstream enforcement of Google rules, and had language that could be interpreted as transferring the license to other models trained on Gemma-generated synthetic data.

Gemma 4 drops all of that for Apache 2.0. No weird clauses, no unilateral updates, no synthetic data transfer concerns. You can use it in commercial products, modify it, redistribute it, and not worry about Google changing the terms later. This puts Gemma 4 in the same licensing bucket as Llama and Mistral, which is where it should have been from the start.

Performance that matters

On the Arena AI text leaderboard, Gemma 31B currently sits at #3 among open models behind GLM-5 and Kimi 2.5. The 26B MoE holds #6. Google claims these models outperform competitors up to 20 times their parameter count. Whether that holds up under independent benchmarking remains to be seen, but the Arena ELO scores are crowd-sourced and hard to game.

The edge models also bring native speech recognition - E2B and E4B can process audio input directly, not just text and images. That is a meaningful upgrade for on-device use cases where piping audio through a separate transcription step adds latency and complexity.

Built for agents

Google is clearly positioning Gemma 4 for the agentic AI wave. All four models support native function calling, structured JSON output, and built-in instructions for common tools and APIs. The code generation quality is reportedly close to what you get from cloud-only services like Claude Code, but it runs entirely offline.

HuggingFace already has full integration across transformers, llama.cpp, MLX, WebGPU, and Rust-based runtimes. Fine-tuning support is available through TRL, Unsloth, and Google own Vertex AI tools.

The 400 million downloads and 100,000 community variants from the Gemma ecosystem suggest there is an eager audience waiting. The Apache 2.0 license removes the last real barrier to adoption.