Google just dropped a bomb on the open source AI world. And for once, it is not a research paper nobody reads.

DeepMind released Gemma 4 today: four model sizes built on the same research that powers Gemini 3, now available to every developer on the planet under an Apache 2.0 license. That last part is the real headline. Google ditched its restrictive custom Gemma license and went fully open source. Developers can modify, redistribute, and commercialize Gemma 4 however they want. No strings attached.

The numbers are absurd. The 31-billion parameter Dense model and the 26-billion Mixture of Experts variant landed third and sixth respectively on Arena AI's text leaderboard, beating models 20 times their size. Let that sink in. A model you can run on a laptop is outperforming systems that require entire data centers.

Google is offering four flavors: 2-billion and 4-billion parameter "Effective" models for edge devices (yes, your phone), plus the 26B MoE and 31B Dense models for heavier workloads. All four handle video and images. The smaller two also process audio and speech. All of them support 140+ languages and can generate code offline. No internet required.

This Is a Direct Shot at Meta

Meta has owned the open source AI conversation for two years. Llama was the default. Llama was the ecosystem. Developers built on Llama because there was nothing else worth building on.

That just changed. Gemma 4 is not just competitive with Llama. It is better in key benchmarks while being dramatically more efficient. And the Apache 2.0 license gives developers more freedom than Meta's custom license ever did. Google is not trying to match Meta's open source strategy. It is trying to replace it entirely.

The timing is brutal for Meta, too. While Mark Zuckerberg sits on an unreleased model codenamed "Avocado" and tries to figure out how to monetize an open ecosystem that everyone else profits from, Google is handing developers a better alternative with fewer restrictions.

The Bigger Picture

Constellation Research flagged something critical: this release comes as the US lags behind China in open source AI. DeepSeek and Qwen have been eating market share while American companies hoard their best models behind paywalls. Gemma 4 is Google's answer to the argument that only China takes open source AI seriously.

While OpenAI and Anthropic guard their weights like state secrets, Google just handed the world a multimodal, multilingual, edge-deployable AI model for free. The open source AI war just went from a two-player game to a three-way fight. And the newcomer showed up with the best hand.

First reported by Engadget and Ars Technica. Gemma 4 is available now on Hugging Face and Google AI Studio.