Google just made its strongest move yet to win the open-weight AI developer market.
Google DeepMind launched Gemma 4 on Wednesday - a family of four open-weight models built from the same research foundation as Gemini 3, available immediately under the Apache 2.0 license. The license change is significant. Previous Gemma releases used a custom license that created commercial uncertainty for developers. Apache 2.0 is the industry standard for commercial permissiveness, giving developers complete control over how they deploy, modify, and build on the models. Hugging Face co-founder Clement Delangue called it "a huge milestone."
What Gemma 4 Includes
The family comes in four sizes targeting different hardware tiers. The Effective 2B and Effective 4B models are optimized for edge devices - smartphones, Raspberry Pi, and Jetson Nano - with native audio input, near-zero latency, and memory usage under 1.5GB for the E2B model. Google worked directly with the Pixel team, Qualcomm, and MediaTek on these builds.
The larger models target developers and enterprises. The 26B Mixture of Experts model activates only 3.8 billion parameters during inference, delivering high tokens-per-second throughput despite its size. The 31B Dense model prioritizes quality and benchmark performance - it currently ranks third among all open models on the Arena AI text leaderboard. Both fit unquantized on a single 80GB Nvidia H100 GPU.
All four models share a common capability set: native image and video processing, support for over 140 languages, function calling, and structured JSON output - meaning developers no longer need to retrofit applications to make models interact with external tools. Context windows reach 128K tokens for the smaller models and 256K for the larger ones.
The Strategic Play
Google is fighting a two-front war in AI. Gemini handles Google's proprietary ecosystem and enterprise cloud business. Gemma is how Google competes for independent developers and organizations that need on-premise deployment, data sovereignty, or the ability to fine-tune models without cloud dependency.
The Gemmaverse already has 400 million downloads and more than 100,000 variants built on previous versions. Real-world examples include INSAIT's Bulgarian-first language model BgGPT and Yale University's Cell2Sentence-Scale cancer research model. These are exactly the use cases Apache 2.0 enables - researchers and sovereign institutions building things that a restrictive commercial license would have blocked.
The competitive timing is pointed. Meta's Llama 4 has been the default open-source choice for most enterprise developers. Gemma 4, with better benchmark performance than previous releases and a genuinely open license, is Google's clearest attempt yet to change that dynamic. For business leaders evaluating AI infrastructure, an open model that runs entirely on local hardware - with no per-token costs, no data leaving the organization, and no vendor lock-in - is worth a serious look.
